Full Width 
Shortcuts 
Sign Up 
Log In 
Articles 
Reading List More from Irrational Exuberance
There’s a lot of excitement about what AI (specifically the latest wave of LLM-anchored AI) can do, and how AI-first companies are different from the prior generations of companies. There are a lot of important and real opportunities at hand, but I find that many of these conversations occur at such an abstract altitude that they’re a bit too abstract. Sort of like saying that your company could be much better if you merely adopted software. That’s certainly true, but it’s not a particularly helpful claim. This post is an attempt to concisely summarize how AI agents work, apply that summary to a handful of real-world use cases for AI, and make the case that the potential of AI agents is equivalent to the potential of this generation of AI. By the end of this writeup, my hope is that you’ll be well-armed to have a concrete discussion about how LLMs and agents could change the shape of your company. How do agents work? At its core, using an LLM is an API call that includes a prompt. For example, you might call Anthropic’s /v1/message with a prompt: How should I adopt LLMs in my company? That prompt is used to fill the LLM’s context window, which conditions the model to generate certain kinds of responses. This is the first important thing that agents can do: use an LLM to evaluate a context window and get a result. Prompt engineering, or context engineering as it’s being called now, is deciding what to put into the context window to best generate the responses you’re looking for. For example, In-Context Learning (ICL) is one form of context engineering, where you supply a bunch of similar examples before asking a question. If I want to determine if a transaction is fraudulent, then I might supply a bunch of prior transactions and whether they were, or were not, fraudulent as ICL examples. Those examples make generating the correct answer more likely. However, composing the perfect context window is very time intensive, benefiting from techniques like metaprompting to improve your context. Indeed, the human (or automation) creating the initial context might not know enough to do a good job of providing relevant context. For example, if you prompt, Who is going to become the next mayor of New York City?, then you are unsuited to include the answer to that question in your prompt. To do that, you would need to already know the answer, which is why you’re asking the question to begin with! This is where we see model chat experiences from OpenAI and Anthropic use web search to pull in context that you likely don’t have. If you ask a question about the new mayor of New York, they use a tool to retrieve web search results, then add the content of those searches to your context window. This is the second important thing that agents can do: use an LLM to suggest tools relevant to the context window, then enrich the context window with the tool’s response. However, it’s important to clarify how “tool usage” actually works. An LLM does not actually call a tool. (You can skim OpenAI’s function calling documentation if you want to see a specific real-world example of this.) Instead there is a five-step process to calling tools that can be a bit counter-intuitive: The program designer that calls the LLM API must also define a set of tools that the LLM is allowed to suggest using. Every API call to the LLM includes that defined set of tools as options that the LLM is allowed to recommend The response from the API call with defined functions is either: Generated text as any other call to an LLM might provide A recommendation to call a specific tool with a specific set of parameters, e.g. an LLM that knows about a get_weather tool, when prompted about the weather in Paris, might return this response: [{ "type": "function_call", "name": "get_weather", "arguments": "{\"location\":\"Paris, France\"}" }] The program that calls the LLM API then decides whether and how to honor that requested tool use. The program might decide to reject the requested tool because it’s been used too frequently recently (e.g. rate limiting), it might check if the associated user has permission to use the tool (e.g. maybe it’s a premium only tool), it might check if the parameters match the user’s role-based permissions as well (e.g. the user can check weather, but only admin users are allowed to check weather in France). If the program does decide to call the tool, it invokes the tool, then calls the LLM API with the output of the tool appended to the prior call’s context window. The important thing about this loop is that the LLM itself can still only do one interesting thing: taking a context window and returning generated text. It is the broader program, which we can start to call an agent at this point, that calls tools and sends the tools’ output to the LLM to generate more context. What’s magical is that LLMs plus tools start to really improve how you can generate context windows. Instead of having to have a very well-defined initial context window, you can use tools to inject relevant context to improve the initial context. This brings us to the third important thing that agents can do: they manage flow control for tool usage. Let’s think about three different scenarios: Flow control via rules has concrete rules about how tools can be used. Some examples: it might only allow a given tool to be used once in a given workflow (or a usage limit of a tool for each user, etc) it might require that a human-in-the-loop approves parameters over a certain value (e.g. refunds more than $100 require human approval) it might run a generated Python program and return the output to analyze a dataset (or provide error messages if it fails) apply a permission system to tool use, restricting who can use which tools and which parameters a given user is able to use (e.g. you can only retrieve your own personal data) a tool to escalate to a human representative can only be called after five back and forths with the LLM agent Flow control via statistics can use statistics to identify and act on abnormal behavior: if the size of a refund is higher than 99% of other refunds for the order size, you might want to escalate to a human if a user has used a tool more than 99% of other users, then you might want to reject usage for the rest of the day it might escalate to a human representative if tool parameters are more similar to prior parameters that required escalation to a human agent LLMs themselves absolutely cannot be trusted. Anytime you rely on an LLM to enforce something important, you will fail. Using agents to manage flow control is the mechanism that makes it possible to build safe, reliable systems with LLMs. Whenever you find yourself dealing with an unreliable LLM-based system, you can always find a way to shift the complexity to a tool to avoid that issue. As an example, if you want to do algebra with an LLM, the solution is not asking the LLM to directly perform algebra, but instead providing a tool capable of algebra to the LLM, and then relying on the LLM to call that tool with the proper parameters. At this point, there is one final important thing that agents do: they are software programs. This means they can do anything software can do to build better context windows to pass on to LLMs for generation. This is an infinite category of tasks, but generally these include: Building general context to add to context window, sometimes thought of as maintaining memory Initiating a workflow based on an incoming ticket in a ticket tracker, customer support system, etc Periodically initiating workflows at a certain time, such as hourly review of incoming tickets Alright, we’ve now summarized what AI agents can do down to four general capabilities. Recapping a bit, those capabilities are: Use an LLM to evaluate a context window and get a result Use an LLM to suggest tools relevant to the context window, then enrich the context window with the tool’s response Manage flow control for tool usage via rules or statistical analysis Agents are software programs, and can do anything other software programs do Armed with these four capabilities, we’ll be able to think about the ways we can, and cannot, apply AI agents to a number of opportunities. Use Case 1: Customer Support Agent One of the first scenarios that people often talk about deploying AI agents is customer support, so let’s start there. A typical customer support process will have multiple tiers of agents who handle increasingly complex customer problems. So let’s set a goal of taking over the easiest tier first, with the goal of moving up tiers over time as we show impact. Our approach might be: Allow tickets (or support chats) to flow into an AI agent Provide a variety of tools to the agent to support: Retrieving information about the user: recent customer support tickets, account history, account state, and so on Escalating to next tier of customer support Refund a purchase (almost certainly implemented as “refund purchase” referencing a specific purchase by the user, rather than “refund amount” to prevent scenarios where the agent can be fooled into refunding too much) Closing the user account on request Include customer support guidelines in the context window, describe customer problems, map those problems to specific tools that should be used to solve the problems Flow control rules that ensure all calls escalate to a human if not resolved within a certain time period, number of back-and-forth exchanges, if they run into an error in the agent, and so on. These rules should be both rules-based and statistics-based, ensuring that gaps in your rules are neither exploitable nor create a terrible customer experience Review agent-customer interactions for quality control, making improvements to the support guidelines provided to AI agents. Initially you would want to review every interaction, then move to interactions that lead to unusual outcomes (e.g. escalations to human) and some degree of random sampling Review hourly, then daily, and then weekly metrics of agent performance Based on your learnings from the metric reviews, you should set baselines for alerts which require more immediate response. For example, if a new topic comes up frequently, it probably means a serious regression in your product or process, and it requires immediate review rather than periodical review. Note that even when you’ve moved “Customer Support to AI agents”, you still have: a tier of human agents dealing with the most complex calls humans reviewing the periodic performance statistics humans performing quality control on AI agent-customer interactions You absolutely can replace each of those downstream steps (reviewing performance statistics, etc) with its own AI agent, but doing that requires going through the development of an AI product for each of those flows. There is a recursive process here, where over time you can eliminate many human components of your business, in exchange for increased fragility as you have more tiers of complexity. The most interesting part of complex systems isn’t how they work, it’s how they fail, and agent-driven systems will fail occasionally, as all systems do, very much including human-driven ones. Applied with care, the above series of actions will work successfully. However, it’s important to recognize that this is building an entire software pipeline, and then learning to operate that software pipeline in production. These are both very doable things, but they are meaningful work, turning customer support leadership into product managers and requiring an engineering team building and operating the customer support agent. Use Case 2: Triaging incoming bug reports When an incident is raised within your company, or when you receive a bug report, the first problem of the day is determining how severe the issue might be. If it’s potentially quite severe, then you want on-call engineers immediately investigating; if it’s certainly not severe, then you want to triage it in a less urgent process of some sort. It’s interesting to think about how an AI agent might support this triaging workflow. The process might work as follows: Pipe all created incidents and all created tickets to this agent for review. Expose these tools to the agent: Open an incident Retrieve current incidents Retrieve recently created tickets Retrieve production metrics Retrieve deployment logs Retrieve feature flag change logs Toggle known-safe feature flags Propose merging an incident with another for human approval Propose merging a ticket with another ticket for human approval Redundant LLM providers for critical workflows. If the LLM provider’s API is unavailable, retry three times over ten seconds, then resort to using a second model provider (e.g. Anthropic first, if unavailable try OpenAI), and then finally create an incident that the triaging mechanism is unavailable. For critical workflows, we can’t simply assume the APIs will be available, because in practice all major providers seem to have monthly availability issues. Merge duplicates. When a ticket comes in, first check ongoing incidents and recently created tickets for potential duplicates. If there is a probable duplicate, suggest merging the ticket or incident with the existing issue and exit the workflow. Assess impact. If production statistics are severely impacted, or if there is a new kind of error in production, then this is likely an issue that merits quick human review. If it’s high priority, open an incident. If it’s low priority, create a ticket. Propose cause. Now that the incident has been sized, switch to analyzing the potential causes of the incident. Look at the code commits in recent deploys and suggest potential issues that might have caused the current error. In some cases this will be obvious (e.g. spiking errors with a traceback of a line of code that changed recently), and in other cases it will only be proximity in time. Apply known-safe feature flags. Establish an allow list of known safe feature flags that the system is allowed to activate itself. For example, if there are expensive features that are safe to disable, it could be allowed to disable them, e.g. restricting paginating through deeper search results when under load might be a reasonable tradeoff between stability and user experience. Defer to humans. At this point, rely on humans to drive incident, or ticket, remediation to completion. Draft initial incident report. If an incident was opened, the agent should draft an initial incident report including the timeline, related changes, and the human activities taken over the course of the incident. This report should then be finalized by the human involved in the incident. Run incident review. Your existing incident review process should take the incident review and determine how to modify your systems, including the triaging agent, to increase reliability over time. Safeguard to reenable feature flags. Since we now have an agent disabling feature flags, we also need to add a periodic check (agent-driven or otherwise) to reenable the “known safe” feature flags if there isn’t an ongoing incident to avoid accidentally disabling them for long periods of time. This is another AI agent that will absolutely work as long as you treat it as a software product. In this case, engineering is likely the product owner, but it will still require thoughtful iteration to improve its behavior over time. Some of the ongoing validation to make this flow work includes: The role of humans in incident response and review will remain significant, merely aided by this agent. This is especially true in the review process, where an agent cannot solve the review process because it’s about actively learning what to change based on the incident. You can make a reasonable argument that an agent could decide what to change and then hand that specification off to another agent to implement it. Even today, you can easily imagine low risk changes (e.g. a copy change) being automatically added to a ticket for human approval. Doing this for more complex, or riskier changes, is possible but requires an extraordinary degree of care and nuance: it is the polar opposite of the idea of “just add agents and things get easy.” Instead, enabling that sort of automation will require immense care in constraining changes to systems that cannot expose unsafe behavior. For example, one startup I know has represented their domain logic in a domain-specific language (DSL) that can be safely generated by an LLM, and are able to represent many customer-specific features solely through that DSL. Expanding the list of known-safe feature flags to make incidents remediable. To do this widely will require enforcing very specific requirements for how software is developed. Even doing this narrowly will require changes to ensure the known-safe feature flags remain safe as software is developed. Periodically reviewing incident statistics over time to ensure mean-time-to-resolution (MTTR) is decreasing. If the agent is truly working, this should decrease. If the agent isn’t driving a reduction in MTTR, then something is rotten in the details of the implementation. Even a very effective agent doesn’t relieve the responsibility of careful system design. Rather, agents are a multiplier on the quality of your system design: done well, agents can make you significantly more effective. Done poorly, they’ll only amplify your problems even more widely. Do AI Agents Represent Entirety of this Generation of AI? If you accept my definition that AI agents are any combination of LLMs and software, then I think it’s true that there’s not much this generation of AI can express that doesn’t fit this definition. I’d readily accept the argument that LLM is too narrow a term, and that perhaps foundational model would be a better term. My sense is that this is a place where frontier definitions and colloquial usage have deviated a bit. Closing thoughts LLMs and agents are powerful mechanisms. I think they will truly change how products are designed and how products work. An entire generation of software makers, and company executives, are in the midst of learning how these tools work. Software isn’t magic, it’s very logical, but what it can accomplish is magical. The same goes for agents and LLMs. The more we can accelerate that learning curve, the better for our industry.
Over the past 19 months, I’ve written Crafting Engineering Strategy, a book on creating engineering strategy. I’ve also been working increasingly with large language models at work. Unsurprisingly, the intersection of those two ideas is a topic that I’ve been thinking about a lot. What, I’ve wondered, is the role of the author, particularly the long-form author, in a world where an increasingly large percentage of writing is intermediated by large language models? One framing I’ve heard somewhat frequently is the view that LLMs are first and foremost a great pillaging of authors’ work. It’s true. They are that. At some point there was a script to let you check which books had been loaded into Meta’s LLaMa, and every book I’d written at that point was included, none of them with my consent. However, I long ago made my peace with plagiarism online, and this strikes me as not particularly different, albeit conducted by larger players. The folks using this writing are going to keep using it beyond the constraints I’d prefer it to be used in, and I’m disinterested in investing my scarce mental energy chasing through digital or legal mazes. Instead, I’ve been thinking about how this transition might go right for authors. My favorite idea that I’ve come up with is the idea of written content as “datapacks” for thinking. Buy someone’s book / “datapack”, then upload it into your LLM, and you can immediately operate almost as if you knew the book’s content. Let’s start with an example. Imagine you want help onboarding as an executive, and you’ve bought a copy of The Engineering Executive’s Primer, you could create a project in Anthropic’s Claude, and upload the LLM-optimized book into your project. Here is what your Claude project might look like. Once you have it set up, you can ask it to help you create your onboarding plan. This guidance makes sense, largely pulled from Your first 90 days as CTO. As always, you can iterate on your initial prompt–including more details you want to include into the plan–along with follow ups to improve the formatting and so on. One interesting thing here, is that I don’t currently have a datapack for The Engineering Executive’s Primer! To solve that, I built one from all my blog posts marked with the “executive” tag. I did that using this script that packages Hugo blog posts, that I generated using this prompt with Claude 3.7 Sonnet. The output of that script gets passed into repomix via: repomix --include "`./scripts/tags.py content executive | paste -d, -s -`" The mess with paste is to turn the multiline output from tags.py into a comma-separated list that repomix knows how to use. This is a really neat pattern, and starts to get at where I see the long-term advantage of writers in the current environment: if you’re a writer and have access to your raw content, you can create a problem-specific datapack to discuss the problem. You can also give that datapack to someone else, or use it to answer their questions. For example, someone asked me a very detailed followup question about a recent blog post. It was a very long question, and I was on a weekend trip. I already had a Claude project setup with the contents of Crafting Engineering Strategy, so I just passed the question verbatim into that project, and sent the answer back to the person who asked it. (I did have to ask Claude to revise the answer once to focus more on what I thought the most important part of the answer was.) This, for what it’s worth, wasn’t a perfect answer, but it’s pretty good. If the question asker had the right datapack, they could have gotten it themselves, without needing me to decide to answer it. However, this post is less worried about the reader than it is about the author. What is our competitive advantage as authors in a future where people are not reading our work? Well, maybe they’re still buying our work in the form of datapacks and such, but it certainly seems likely that book sales, like blog traffic, will be impacted negatively. In trade, it’s now possible for machines to understand our thinking that we’ve recorded down into words over time. There’s a running joke in my executive learning circle that I’ve written a blog post on every topic that comes up, and that’s kind of true. That means that I am on the cusp of the opportunity to uniquely scale myself by connecting “intelligence on demand for a few cents” with the written details of my thinking built over the past two decades of being a writer who operates. The tools that exist today are not quite there yet, although a combination of selling datapacks like the one for Crafting Engineering Strategy and tools like Claude’s projects are a good start. There are many ways the exact details might come together, but I’m optimistic that writing will become more powerful rather than less in this new world, even if the particular formats change. (For what it’s worth, I don’t think human readers are going away either.) If you’re interested in the fully fleshed out version of this idea, starting here you can read the full AI Companion to Crafting Engineering Strategy. The datapack will be available via O’Reilly in the next few months. If you’re an existing O’Reilly author who’s skepical of this idea, don’t worry: I worked with them to sign a custom contract, this usage–as best I understood it, although I am not a lawyer and am not providing legal advice–is outside of the scope of the default contract I signed with my prior book, and presumably most others’ contracts as well.
Since 2020, I’ve been working on my desk setup, and I think I finally have it mostly pulled together at this point. I don’t really think my desk setup is very novel, and I’m sure there are better ways to pull it together, but I will say that it finally works the way I want since I added the CalDigit TS5 Plus, which has been a long time coming. My requirements for my desk are: Has support for 2-3 Mac laptops Has support for a Windows gaming desktop with a dedicated GPU Has a dedicated microphone Has good enough lighting Is not too messy I can switch between any laptop and desktop with a single Thunderbolt cable Historically the issue here has been the final requirement, where switching required moving two cables–a Thunderbolt and a cable for the dedicated graphics card–but with my new dock this finally works with just one cable. The equipment shown here, and my brief review of each piece, is: UPLIFT v2 Standing Desk – is the standing desk I use. I both have a lot of stuff on my desk, and also want my desk to feel minimal, so I opted for the 72" x 30" verison. At the time I ordered it in 2020, the only option shipping quickly was the bamboo finish, so that’s what I got. CalDigit TS5 Plus Dock – this was the missing component that has three Thunderbolt ports and a DisplayPort. I have the external graphics card directly connected to the DisplayPort, and then move the Thunderbolt port from computer to computer to change which one is active. It also has enough USB-A ports to connect the adapters for my wireless keyboard and mouse, to avoid needing to pair them across computers which would create friction in switching computers. Apple Studio Display – I experimented with dedicated speakers and video camera, but for me having them built into the monitor was helpful to reduce the number of things on my desk. The Studio Display’s monitor, speakers and video camera are all solidly good enough for my purposes: I’m sure I could get better on each dimension, but in practice I never think about this and don’t find any issues with them. On the other hand, while I was initially hopeful that I could also get rid of my microphone, the microphone quality just wasn’t that good for me, as I spend a lot of time on video conferences and recording podcasts, etc. Beelink GTi Ultra & EX Pro Docking Station – are my Windows mini desktop and dock which allows mounting an external GPU to the mini desktop. Beelink itself is slightly aggrevating because as best I can tell they’ve done something quite odd in terms of custom patching Windows 11, but ultimately it’s worked well for me as a dedicated gaming machine, and the build quality and size profile are both just fantastic. MSI Gaming RTX 4070 Ti Super 16G Graphics Card – I bought this earlier this year, looking for something that was in stock, and was good enough that it would last me a generation or two of graphics card upgrades without shelling out a truly massive amount for a 50XX edition (some of which don’t seem to be upgrades on the 40XX predecesors anyway). Hexcal Studio – this is the workstation / monitor stand / cable management system, with lighting and so on. I ultimately do like this, but it’s not perfect, e.g. my Qi charger technically works but provides such bad charging speeds that it effectively doesn’t work. It’s definitely too expensive for something that doesn’t entirely work, so I can’t really recommend it, although now that I’ve paid for it, I wouldn’t bother replacing it either. Audio-Technica AT2020USB Cardioid Condenser USB Microphone – this is the microphone I’ve been using for six years, and it’s really quite good and cost something like $120 at the time. It’s discontinued now, but presumably there’s a more modern version somewhere. I have it mounted on this boom arm. LUME CUBE Edge 2.0 LED Desk Lamp – I have two of these for lighting during recordings. I don’t actually like using them very much, I just hate looking into lights, but I do use them periodically when I want to make sure lighting is actually correct. Logitech MX Keys Advanced Wireless Illuminated Keyboard for Mac – this keyboard works well for me, and has a USB-C so I can use a single powered USB-C cable from the Hexcal to charge my keyboard, my mouse, my phone, and my headphones. Logitech MX Master 3S Wireless Mouse – I’ve been using variations of this mouse for a long time, I specifically bought this version a year or two ago to standardize all charging ports on USB-C. Laptop stand – I’m not actually sure where I got this laptop stand from, it might have been Etsy. I found it relatively hard to find stands that support three laptops rather than just two. Before finding this one, I used this two-laptop stand which is fine. Laptops – these are my personal and work Macbooks. Here’s a slightly closer look at the left side of the desk. At this point, I really have nothing left that I’m upset about with my setup, and I can’t imagine changing this again in the next few years. As a bonus, my office has a handful of pieces of “professional art” that represent things I am proud of. From left to right, it’s the cover of An Elegant Puzzle, a map of San Francisco drawn exclusively from Uber trip data on the night of Halloween 2014, and then the cover of The Engineering Executive’s Primer. It’s probably a bit vain, but I like to remember some of the accomplishments.
Today’s my last day at Carta, where I got the chance to serve as their CTO for the past two years. I’ve learned so much working there, and I wanted to end my chapter there by collecting my thoughts on what I learned. (I am heading somewhere, and will share news in a week or two after firming up the communication plan with my new team there.) The most important things I learned at Carta were: Working in the details – if you took a critical lens towards my historical leadership style, I think the biggest issue you’d point at is my being too comfortable operating at a high level of abstraction. Utilizing the expertise of others to fill in your gaps is a valuable skill, but–like any single approach–it’s limiting when utilized too frequently. One of the strengths of Carta’s “house leadership style” is expecting leaders to go deep into the details to get informed and push pace. What I practiced there turned into the pieces on strategy testing and developing domain expertise. Refining my approach to engineering strategy – over the past 18 months, I’ve written a book on engineering strategy (posts are all in #eng-strategy-book), with initial chapters coming available for early release with O’Reilly next month. Fingers crossed, the book will be released in approximately October. Coming into Carta, I already had much of my core thesis about how to do engineering strategy, but Carta gave me a number of complex projects to practice on, and excellent people to practice with: thank you to Dan, Shawna and Vogl in particular! More on this project in the next few weeks. Extract the kernel – everywhere I’ve ever worked, teams have struggled understanding executives. In every case, the executives could be clearer, but it’s not particularly interesting to frame these problems as something the executives need to fix. Sure, that’s true they could communicate better, but that framing makes you powerless, when you have a great deal of power to understand confusing communication. After all, even good communicators communicate poorly sometimes. Meaningfully adopting LLMs – a year ago I wrote up notes on adopting LLMs in your products, based on what we’d learned so far. Since then, we’ve learned a lot more, and LLMs themselves have significantly improved. Carta has been using LLMs in real, business-impacting workflows for over a year. That’s continuing to expand into solving more complex internal workflows, and even more interestingly into creating net-new product capabilities that ought to roll out more widely in the next few months (currently released to small beta groups). This is the first major technology transition that I’ve experienced in a senior leadership role (since I was earlier in my career when mobile internet transitioned from novelty to commodity). The immense pressure to adopt faster, combined with the immense uncertainty if it’s a meaningful change or a brief blip was a lot of fun, and was the inspiration for this strategy document around LLM adoption. Multi-dimensional tradeoffs – a phrase that Henry Ward uses frequent is that “everyone’s right, just at a different altitude.” That idea resonates with me, and meshes well with the ideas of multi-dimensional tradeoffs and layers of context that I find improve decision making for folks in roles that require making numerous, complex decisions. Working at Carta, these ideas formalized from something I intuited into something I could explain clearly. Navigators – I think our most successful engineering strategy at Carta was rolling out the Navigator program, which ensured senior-most engineers had context and direct representation, rather than relying exclusively on indirect representation via engineering management. Carta’s engineering managers are excellent, but there’s always something lost as discussions extend across layers. The Navigator program probably isn’t a perfect fit for particularly small companies, but I think any company with more than 100-150 engineers would benefit from something along these lines. How to create software quality – I’ve evolved my thinking about software quality quite a bit over time, but Carta was particularly helpful in distinguishing why some pieces of software are so hard to build despite having little-to-no scale from a data or concurrency perspective. These systems, which I label as “high essential complexity”, deserve more credit for their complexity, even if they have little in the way of complexity from infrastructure scaling. Shaping eng org costs – a few years ago, I wrote about my mental model for managing infrastructure costs. At Carta, I got to refine my thinking about engineering salary costs, with most of those ideas getting incorporated in the Navigating Private Equity ownership strategy, and the eng org seniority mix model. The three biggest levers are (1) “N-1 backfills”, (2) requiring a business rationale for promotions into senior-most levels, and (3) shifting hiring into cost efficient hiring regions. None of these are the sort of inspiring topics that excite folks, but they are all essential to the long term stability of your organization. Explaining engineering costs to boards/execs – Similarly, I finally have a clear perspective on how to represent R&D investment to boards in the same language that they speak in, which I wrote up here, and know how to do it quickly without relying on any manually curated internal datasets. Lots of smaller stuff, like the no wrong doors policy for routing colleagues to appropriate channels, how to request headcount in a way that is convincing to executives, Act Two rationales for how people’s motivations evolve over the course of long careers (and my own personal career mission to advance the industry, why friction isn’t velocity even though many folks act like it is. I’ve also learned quite a bit about venture capital, fund administration, cap tables, non-social network products, operating a multi-business line company, and various operating models. Figuring out how to sanitize those learnings to share the interesting tidbits without leaking internal details is a bit too painful, so I’m omitting them for now. Maybe some will be shareable in four or five years after my context goes sufficiently stale. As a closing thought, I just want to say how much I’ve appreciated the folks I’ve gotten to work with at Carta. From the executive team (Ali, April, Charly, Davis, Henry, Jeff, Nicole, Vrushali) to my directs (Adi, Ciera, Dan, Dave, Jasmine, Javier, Jayesh, Karen, Madhuri, Sam, Shawna) to the navigators (there’s a bunch of y’all). The people truly are always the best part, and that was certainly true at Carta.
Back in 2018, I wrote lethain/systems as a domain-specific language for writing runnable systems models, and introduced it with this blog post modeling a hiring funnel. While it’s far from a perfect system, I’ve gotten a lot of value out of it over the last seven years, because it allows me to maintain systems models in version control. As I’ve been playing with writing Model Context Protocol (MCP) servers, one I’ve been thinking about frequently is one to help writing systems syntax, and I finally put that together in the lethain/systems-mcp repository. More detailed installation and usage instructions are in the GitHub repository, so I’ll just share a couple of screenshots and comments here. Starting with the load_systems_documentation tool which loads a copy of lethain/systems/README.md and a file with example systems into the context window. The biggest challenge of properly writing DSLs with an LLM is providing enough in-context learning (ICL) examples, and I think the idea of providing tools that are specifically designed to provide that context is a very interesting idea. Eventually I imagine there will be generalized tools for this, e.g. a search index of the best ICL examples for a wide variety of DSLs. Until then, my guess is that this sort of tool is particularly valuable. The second tool is run_systems_model which passes the DSL (and an optional parameter for number of rounds) to the tool and then returns the result. I experimented with interface design here, initially trying to return a rendered chart of the results, but ultimately even multi-modal models are just much better at working with text than with images. This meant that I had the best results returning JSON of the results and then having the LLM build a tool for interacting with the results. Altogether, a fun little experiment, and another confirmation in my mind that the most interesting part of designing MCPs today is deciding where to introduce and eliminate complexity from the LLM. Introduce too little and the tool lacks power; eliminate too little and the combination rarely works.
More in programming
<![CDATA[It has been a year since I set up my System76 Merkaat with Linux Mint. In July of 2024 I migrated from ChromeOS and the Merkaat has been my daily driver on the desktop. A year later I have nothing major to report, which is the point. Despite the occasional unplanned reinstallation I have been enjoying the stability of Linux and just using the PC. This stability finally enabled me to burn bridges with mainstream operating systems and fully embrace Linux and open systems. I'm ready to handle the worst and get back to work. Just a few years ago the frustration of troubleshooting a broken system would have made me seriously consider the switch to a proprietary solution. But a year of regular use, with an ordinary mix of quiet moments and glitches, gave me the confidence to stop worrying and learn to love Linux. linux a href="https://remark.as/p/journal.paoloamoroso.com/my-first-year-since-coming-back-to-linux"Discuss.../a Email | Reply @amoroso@oldbytes.space !--emailsub--]]>
There’s a lot of excitement about what AI (specifically the latest wave of LLM-anchored AI) can do, and how AI-first companies are different from the prior generations of companies. There are a lot of important and real opportunities at hand, but I find that many of these conversations occur at such an abstract altitude that they’re a bit too abstract. Sort of like saying that your company could be much better if you merely adopted software. That’s certainly true, but it’s not a particularly helpful claim. This post is an attempt to concisely summarize how AI agents work, apply that summary to a handful of real-world use cases for AI, and make the case that the potential of AI agents is equivalent to the potential of this generation of AI. By the end of this writeup, my hope is that you’ll be well-armed to have a concrete discussion about how LLMs and agents could change the shape of your company. How do agents work? At its core, using an LLM is an API call that includes a prompt. For example, you might call Anthropic’s /v1/message with a prompt: How should I adopt LLMs in my company? That prompt is used to fill the LLM’s context window, which conditions the model to generate certain kinds of responses. This is the first important thing that agents can do: use an LLM to evaluate a context window and get a result. Prompt engineering, or context engineering as it’s being called now, is deciding what to put into the context window to best generate the responses you’re looking for. For example, In-Context Learning (ICL) is one form of context engineering, where you supply a bunch of similar examples before asking a question. If I want to determine if a transaction is fraudulent, then I might supply a bunch of prior transactions and whether they were, or were not, fraudulent as ICL examples. Those examples make generating the correct answer more likely. However, composing the perfect context window is very time intensive, benefiting from techniques like metaprompting to improve your context. Indeed, the human (or automation) creating the initial context might not know enough to do a good job of providing relevant context. For example, if you prompt, Who is going to become the next mayor of New York City?, then you are unsuited to include the answer to that question in your prompt. To do that, you would need to already know the answer, which is why you’re asking the question to begin with! This is where we see model chat experiences from OpenAI and Anthropic use web search to pull in context that you likely don’t have. If you ask a question about the new mayor of New York, they use a tool to retrieve web search results, then add the content of those searches to your context window. This is the second important thing that agents can do: use an LLM to suggest tools relevant to the context window, then enrich the context window with the tool’s response. However, it’s important to clarify how “tool usage” actually works. An LLM does not actually call a tool. (You can skim OpenAI’s function calling documentation if you want to see a specific real-world example of this.) Instead there is a five-step process to calling tools that can be a bit counter-intuitive: The program designer that calls the LLM API must also define a set of tools that the LLM is allowed to suggest using. Every API call to the LLM includes that defined set of tools as options that the LLM is allowed to recommend The response from the API call with defined functions is either: Generated text as any other call to an LLM might provide A recommendation to call a specific tool with a specific set of parameters, e.g. an LLM that knows about a get_weather tool, when prompted about the weather in Paris, might return this response: [{ "type": "function_call", "name": "get_weather", "arguments": "{\"location\":\"Paris, France\"}" }] The program that calls the LLM API then decides whether and how to honor that requested tool use. The program might decide to reject the requested tool because it’s been used too frequently recently (e.g. rate limiting), it might check if the associated user has permission to use the tool (e.g. maybe it’s a premium only tool), it might check if the parameters match the user’s role-based permissions as well (e.g. the user can check weather, but only admin users are allowed to check weather in France). If the program does decide to call the tool, it invokes the tool, then calls the LLM API with the output of the tool appended to the prior call’s context window. The important thing about this loop is that the LLM itself can still only do one interesting thing: taking a context window and returning generated text. It is the broader program, which we can start to call an agent at this point, that calls tools and sends the tools’ output to the LLM to generate more context. What’s magical is that LLMs plus tools start to really improve how you can generate context windows. Instead of having to have a very well-defined initial context window, you can use tools to inject relevant context to improve the initial context. This brings us to the third important thing that agents can do: they manage flow control for tool usage. Let’s think about three different scenarios: Flow control via rules has concrete rules about how tools can be used. Some examples: it might only allow a given tool to be used once in a given workflow (or a usage limit of a tool for each user, etc) it might require that a human-in-the-loop approves parameters over a certain value (e.g. refunds more than $100 require human approval) it might run a generated Python program and return the output to analyze a dataset (or provide error messages if it fails) apply a permission system to tool use, restricting who can use which tools and which parameters a given user is able to use (e.g. you can only retrieve your own personal data) a tool to escalate to a human representative can only be called after five back and forths with the LLM agent Flow control via statistics can use statistics to identify and act on abnormal behavior: if the size of a refund is higher than 99% of other refunds for the order size, you might want to escalate to a human if a user has used a tool more than 99% of other users, then you might want to reject usage for the rest of the day it might escalate to a human representative if tool parameters are more similar to prior parameters that required escalation to a human agent LLMs themselves absolutely cannot be trusted. Anytime you rely on an LLM to enforce something important, you will fail. Using agents to manage flow control is the mechanism that makes it possible to build safe, reliable systems with LLMs. Whenever you find yourself dealing with an unreliable LLM-based system, you can always find a way to shift the complexity to a tool to avoid that issue. As an example, if you want to do algebra with an LLM, the solution is not asking the LLM to directly perform algebra, but instead providing a tool capable of algebra to the LLM, and then relying on the LLM to call that tool with the proper parameters. At this point, there is one final important thing that agents do: they are software programs. This means they can do anything software can do to build better context windows to pass on to LLMs for generation. This is an infinite category of tasks, but generally these include: Building general context to add to context window, sometimes thought of as maintaining memory Initiating a workflow based on an incoming ticket in a ticket tracker, customer support system, etc Periodically initiating workflows at a certain time, such as hourly review of incoming tickets Alright, we’ve now summarized what AI agents can do down to four general capabilities. Recapping a bit, those capabilities are: Use an LLM to evaluate a context window and get a result Use an LLM to suggest tools relevant to the context window, then enrich the context window with the tool’s response Manage flow control for tool usage via rules or statistical analysis Agents are software programs, and can do anything other software programs do Armed with these four capabilities, we’ll be able to think about the ways we can, and cannot, apply AI agents to a number of opportunities. Use Case 1: Customer Support Agent One of the first scenarios that people often talk about deploying AI agents is customer support, so let’s start there. A typical customer support process will have multiple tiers of agents who handle increasingly complex customer problems. So let’s set a goal of taking over the easiest tier first, with the goal of moving up tiers over time as we show impact. Our approach might be: Allow tickets (or support chats) to flow into an AI agent Provide a variety of tools to the agent to support: Retrieving information about the user: recent customer support tickets, account history, account state, and so on Escalating to next tier of customer support Refund a purchase (almost certainly implemented as “refund purchase” referencing a specific purchase by the user, rather than “refund amount” to prevent scenarios where the agent can be fooled into refunding too much) Closing the user account on request Include customer support guidelines in the context window, describe customer problems, map those problems to specific tools that should be used to solve the problems Flow control rules that ensure all calls escalate to a human if not resolved within a certain time period, number of back-and-forth exchanges, if they run into an error in the agent, and so on. These rules should be both rules-based and statistics-based, ensuring that gaps in your rules are neither exploitable nor create a terrible customer experience Review agent-customer interactions for quality control, making improvements to the support guidelines provided to AI agents. Initially you would want to review every interaction, then move to interactions that lead to unusual outcomes (e.g. escalations to human) and some degree of random sampling Review hourly, then daily, and then weekly metrics of agent performance Based on your learnings from the metric reviews, you should set baselines for alerts which require more immediate response. For example, if a new topic comes up frequently, it probably means a serious regression in your product or process, and it requires immediate review rather than periodical review. Note that even when you’ve moved “Customer Support to AI agents”, you still have: a tier of human agents dealing with the most complex calls humans reviewing the periodic performance statistics humans performing quality control on AI agent-customer interactions You absolutely can replace each of those downstream steps (reviewing performance statistics, etc) with its own AI agent, but doing that requires going through the development of an AI product for each of those flows. There is a recursive process here, where over time you can eliminate many human components of your business, in exchange for increased fragility as you have more tiers of complexity. The most interesting part of complex systems isn’t how they work, it’s how they fail, and agent-driven systems will fail occasionally, as all systems do, very much including human-driven ones. Applied with care, the above series of actions will work successfully. However, it’s important to recognize that this is building an entire software pipeline, and then learning to operate that software pipeline in production. These are both very doable things, but they are meaningful work, turning customer support leadership into product managers and requiring an engineering team building and operating the customer support agent. Use Case 2: Triaging incoming bug reports When an incident is raised within your company, or when you receive a bug report, the first problem of the day is determining how severe the issue might be. If it’s potentially quite severe, then you want on-call engineers immediately investigating; if it’s certainly not severe, then you want to triage it in a less urgent process of some sort. It’s interesting to think about how an AI agent might support this triaging workflow. The process might work as follows: Pipe all created incidents and all created tickets to this agent for review. Expose these tools to the agent: Open an incident Retrieve current incidents Retrieve recently created tickets Retrieve production metrics Retrieve deployment logs Retrieve feature flag change logs Toggle known-safe feature flags Propose merging an incident with another for human approval Propose merging a ticket with another ticket for human approval Redundant LLM providers for critical workflows. If the LLM provider’s API is unavailable, retry three times over ten seconds, then resort to using a second model provider (e.g. Anthropic first, if unavailable try OpenAI), and then finally create an incident that the triaging mechanism is unavailable. For critical workflows, we can’t simply assume the APIs will be available, because in practice all major providers seem to have monthly availability issues. Merge duplicates. When a ticket comes in, first check ongoing incidents and recently created tickets for potential duplicates. If there is a probable duplicate, suggest merging the ticket or incident with the existing issue and exit the workflow. Assess impact. If production statistics are severely impacted, or if there is a new kind of error in production, then this is likely an issue that merits quick human review. If it’s high priority, open an incident. If it’s low priority, create a ticket. Propose cause. Now that the incident has been sized, switch to analyzing the potential causes of the incident. Look at the code commits in recent deploys and suggest potential issues that might have caused the current error. In some cases this will be obvious (e.g. spiking errors with a traceback of a line of code that changed recently), and in other cases it will only be proximity in time. Apply known-safe feature flags. Establish an allow list of known safe feature flags that the system is allowed to activate itself. For example, if there are expensive features that are safe to disable, it could be allowed to disable them, e.g. restricting paginating through deeper search results when under load might be a reasonable tradeoff between stability and user experience. Defer to humans. At this point, rely on humans to drive incident, or ticket, remediation to completion. Draft initial incident report. If an incident was opened, the agent should draft an initial incident report including the timeline, related changes, and the human activities taken over the course of the incident. This report should then be finalized by the human involved in the incident. Run incident review. Your existing incident review process should take the incident review and determine how to modify your systems, including the triaging agent, to increase reliability over time. Safeguard to reenable feature flags. Since we now have an agent disabling feature flags, we also need to add a periodic check (agent-driven or otherwise) to reenable the “known safe” feature flags if there isn’t an ongoing incident to avoid accidentally disabling them for long periods of time. This is another AI agent that will absolutely work as long as you treat it as a software product. In this case, engineering is likely the product owner, but it will still require thoughtful iteration to improve its behavior over time. Some of the ongoing validation to make this flow work includes: The role of humans in incident response and review will remain significant, merely aided by this agent. This is especially true in the review process, where an agent cannot solve the review process because it’s about actively learning what to change based on the incident. You can make a reasonable argument that an agent could decide what to change and then hand that specification off to another agent to implement it. Even today, you can easily imagine low risk changes (e.g. a copy change) being automatically added to a ticket for human approval. Doing this for more complex, or riskier changes, is possible but requires an extraordinary degree of care and nuance: it is the polar opposite of the idea of “just add agents and things get easy.” Instead, enabling that sort of automation will require immense care in constraining changes to systems that cannot expose unsafe behavior. For example, one startup I know has represented their domain logic in a domain-specific language (DSL) that can be safely generated by an LLM, and are able to represent many customer-specific features solely through that DSL. Expanding the list of known-safe feature flags to make incidents remediable. To do this widely will require enforcing very specific requirements for how software is developed. Even doing this narrowly will require changes to ensure the known-safe feature flags remain safe as software is developed. Periodically reviewing incident statistics over time to ensure mean-time-to-resolution (MTTR) is decreasing. If the agent is truly working, this should decrease. If the agent isn’t driving a reduction in MTTR, then something is rotten in the details of the implementation. Even a very effective agent doesn’t relieve the responsibility of careful system design. Rather, agents are a multiplier on the quality of your system design: done well, agents can make you significantly more effective. Done poorly, they’ll only amplify your problems even more widely. Do AI Agents Represent Entirety of this Generation of AI? If you accept my definition that AI agents are any combination of LLMs and software, then I think it’s true that there’s not much this generation of AI can express that doesn’t fit this definition. I’d readily accept the argument that LLM is too narrow a term, and that perhaps foundational model would be a better term. My sense is that this is a place where frontier definitions and colloquial usage have deviated a bit. Closing thoughts LLMs and agents are powerful mechanisms. I think they will truly change how products are designed and how products work. An entire generation of software makers, and company executives, are in the midst of learning how these tools work. Software isn’t magic, it’s very logical, but what it can accomplish is magical. The same goes for agents and LLMs. The more we can accelerate that learning curve, the better for our industry.
The mystery In the previous article, I briefly mentioned a slight difference between the ESP-Prog and the reproduced circuit, when it comes to EN: Focusing on EN, it looks like the voltage level goes back to 3.3V much faster on the ESP-Prog than on the breadboard circuit. The grid is horizontally spaced at 2ms, so … Continue reading Overanalyzing a minor quirk of Espressif’s reset circuit → The post Overanalyzing a minor quirk of Espressif’s reset circuit appeared first on Quentin Santos.
After 18 months of living in Karuizawa, a resort town about an hour away from Tokyo via the Shinkansen, I have bought a house here. This article describes my experience of purchasing a house, and contains tips that are useful both if you’re considering buying in Karuizawa specifically, and in Japan more generally. In this article, I’ll cover: My personal journey to Karuizawa Why you might choose to live in Karuizawa My process of buying a house Tips for potential homebuyers in Japan From Tokyo to Karuizawa: my own journey In April 2020 I was living in central Tokyo. Three days after my one-year-old son entered daycare, a state of emergency was declared because of COVID-19. The daycare was closed, so my wife and I took turns looking after him while working remotely from our apartment. The small park next to our apartment was also closed due to the state of emergency, and so my time with my son included such fun activities as walking up and down the stairs, and looking at ants in the parking lot. After a week or two, we’d had enough. We moved in with my in-laws, who were living in a house in a small commuter town in Saitama. Our lives improved dramatically. The nearby parks were not closed. The grandparents could help look after my son. I could enjoy excellent cycling in the nearby mountains. We kept our apartment in Tokyo, as we didn’t know how long we’d stay, but let my brother-in-law use it. It was more spacious than his own, and gave him a better place to work remotely from. Leaving Tokyo for good In June of 2020 the state of emergency was lifted, yet we decided not to move back to Tokyo. Before having a kid, Tokyo had been a wonderful place to live. There was always something new to try, and opportunities to make new professional connections abounded. But as a father, my lifestyle had changed dramatically, even before COVID. No longer was I attending multiple events per week, but at most one per month. I basically stopped drinking alcohol, as maximizing the quality of what little sleep I could get was too important. With a stroller, getting around by train was no longer easy. My life had become much more routine: excitement came from watching my son grow, not my own activities. Though we’d enjoyed living in an area that was suburban bordering on rural, we didn’t want to live with my in-laws indefinitely. Still, seeing how useful it was to have them around with a small child, we decided to look for a house to rent nearby. We found a relatively modern one in the next town over, about a 10-minute drive from my in-laws. The rent was about half of what our Tokyo apartment’s had been, yet this house was twice as big, and even had a small garden. Living there was a wonderful change of pace. When my second child was born a year later, having the in-laws nearby became even more helpful. But as the COVID restrictions began rolling back, we started to think about where we wanted to settle down permanently. Choosing Karuizawa There were some definite downsides to the town we were living in. Practically all our neighbours were retirees. There were few restaurants that were worth going to. The equipment at the playgrounds was falling apart, and there was no sign it would be replaced. The summers were brutally hot, making it unpleasant to be outside. There wasn’t anything for kids to do inside. We’d visited Karuizawa before, and it seemed like it had the access to nature we appreciated, while also offering more options for dining out and other cultural activities. The cooler climate was also a big motivating factor. So in June 2022, we started looking at places. Most of the rental options were incredibly expensive seasonal properties, and so buying a house seemed like the most realistic option. But the market moved extremely quickly, and any property that seemed like a good deal was snatched up before we even had a chance to look at it. After about half a year of searching, we gave up. We then considered buying a house in Gotenba, a town at the base of Mount Fuji. We’d visited there in the past, and though it didn’t seem quite as nice an option as Karuizawa, we were able to find a decent house that was cheap enough that buying it was a relatively small risk. We placed a lowball offer on the property, and were rejected. Frustrated, I took a look at the rental options for Karuizawa again. This time we saw two houses for rent that were good candidates! We heard from the real estate agent that one of the properties already had someone coming to view the next afternoon, and so we went the next morning. Visiting the house, it seemed like exactly what we were looking for. We immediately put in a rental application, and moved in the next month. Why Karuizawa? I’ve already covered some of the reasons why we chose Karuizawa, but I’ll go into more detail about them, along with other unexpected good points. Mild summers According to historical data, the average temperature in Karuizawa in August is 20.8°C. While daytime temperatures get hotter than that, you’re unlikely to risk heat stroke by being active outside, something that isn’t true for much of Japan. That being said, climate change is inescapable, and summers in Karuizawa are getting hotter than historical data would suggest. This is an issue because while it cools down at night, many of the houses are built only with the cold winters in mind. For instance, despite the house I rented being built in 2019, it only had a single AC unit. This meant that many days in July and August were unbearably hot inside, and I had trouble sleeping at night as it wouldn’t cool down until well past midnight. Cold but cheerful winters While summers are cooler, this also means winters are cold. The average temperature in January is -3.3°C. However, much like Tokyo, winters here tend to be quite sunny, and there have only been a few heavy snowfalls per year while I’ve lived here. That’s enough that you can enjoy the snow without having to worry about shovelling it daily. Proximity to Tokyo The Shinkansen makes it quicker for me to get into central Tokyo than when I was living in suburban Saitama. It’s about a 70-minute ride from Karuizawa station to Tokyo station. While this is not something I’d want to do daily, I now typically go into Tokyo only a couple of times per month, which is reasonable. Greenery and nature As a resort community, Karuizawa places a lot of emphasis on greenery. Throughout the town, there are plenty of areas where residential and commercial spaces mingle with forest. Many houses have nice gardens. Shade exists everywhere. Wilder nature is close by, too. While you can do day hikes from Tokyo, they involve an hour plus on the train, followed by a climb where you’re usually surrounded by other people. In Karuizawa, I can walk 15 minutes to a trailhead where I’ll see at most a couple of groups per hour. Cheaper than Tokyo, but not significantly so Money spent on housing goes a lot further in Karuizawa than in Tokyo. For the same price as a used Tokyo apartment, you can get a nice used house on a relatively large plot of land in Karuizawa. However, it’s hard to find truly affordable housing. For starters, Karuizawa’s plots tend to be large by Japanese standards, so even if the per-unit price of the land is cheaper, the overall cost isn’t so cheap. For instance, a centrally located 100 tsubo (330 m²) plot in Nakakaruizawa, which is on the small side of what you can find, will currently sell for at least 30 million yen. It may be possible to get a deal on a more remote piece of land, but true steals are hard to come by. Dining out is another area where it’s easy to spend lots of money. Prices are on par with Tokyo, and there are few cheap options, with Sukiya, McDonald’s, and a couple of ramen shops being practically the only sub-1,000 yen choices. Groceries, at least, are cheaper. While the Tsuruya grocery store does sell a lot of high-end items, the basics like milk, eggs, vegetables, and chicken are significantly cheaper than in Tokyo or even suburban Saitama. All this means that many residents of Karuizawa are wealthy compared to Japan’s overall population. Sometimes that makes them take for granted that others might struggle financially. For instance, when my son graduated from public daycare, the PTA organized a graduation party. We paid 16,000 yen for myself, my wife, and two children to attend. Since that’s equivalent to two eight-hour days at minimum wage, I think it was too expensive. I noticed that some of the children didn’t attend, while others came with just one parent, perhaps because of the prohibitive cost. Liquid housing market While we were living in suburban Saitama, we considered buying a house, and even looked at a couple of places. It was very affordable, with even mansions (in the traditional English usage of the word) being cheaper than Tokyo mansions (in the Japanese sense). But the reason they were so cheap was because there wasn’t a secondary market for them. So even if we could get a good deal, we were worried that should we ever want to sell it again, it might not even be possible to give it away. Karuizawa’s real estate market moves fast. This sucks as a buyer but is great for sellers. Furthermore, because it’s perceived as a place to have a second home by wealthy Tokyoites, houses that are by our standards very nice are nowhere near the top of the market. This reduces the risk of buying a property and not being able to sell it later. Great public facilities If you have young children, Karuizawa offers many free or cheap public facilities. This includes nice playgrounds, children’s halls stocked with toys and activities, a beautiful library with a decent selection of English picture books, and a recreation centre with a gym, training room, pool, indoor and outdoor skating rinks, and a curling rink (built for 1998 Winter Olympics). A welcoming and international community Some areas of Japan have a reputation for being insular and not friendly toward newcomers. Perhaps because much of Karuizawa’s population has moved here from elsewhere, the people I’ve met have been quite welcoming. I’m also always happy to meet new residents, and pass on the helpful tips I’ve picked up. In addition, I’ve discovered that many residents have some sort of international background. While non-Japanese residents are very much the minority, a large percentage of the Japanese parents I’ve spoken with have either studied or worked abroad. There is also an International Association of Karuizawa, whose members include many long-term non-Japanese residents. Their Facebook group makes it easy to ask for advice in English. You can get around without a car Most days I don’t use a car, and get around by walking or bicycling. The lack of heavy snow makes it possible to do this year round (though I’m one of the few people who still uses a bicycle when it is -10°C out). Mostly I’m able to get around without a car because I live in a relatively central area, so it’s not feasible for every resident. It can be quite helpful, though, as traffic becomes absolutely horrible during the peak seasons. For instance, the house I was renting was a kilometer away from the one I bought. It was less than a 5-minute drive normally, but during Golden Week it took 30 minutes. That being said, it’s practically required to use a car sometimes. For instance, there’s a great pediatrician, but they’re only accessible by car. Similarly, I use our car to take my kids to the pool, pick up things from the home centre, and so on. My process for buying a house After a year of renting in Karuizawa, we decided we wanted to continue living there indefinitely. Continuing to rent the same house wasn’t an option: the owners were temporarily abroad and so we had a fixed-term contract. While we could have looked for another place to rent, we figured we’d be settling down, and so wanted to buy a house. Starting the search On September 1st, 2024, we started our search and immediately found our first candidate: a recently-built house near Asama Fureai Park. That park is the nicest one for young children in Karuizawa, as it has a good playground but is never too crowded. The property was listed with Royal Resort, which has the most listings for Karuizawa. Rather than contacting Royal Resort directly, though, we asked Masaru Takeyama of Resort Innovation to do so on our behalf. In Japan, there is a buyer’s agent and seller’s agent. Often the seller’s agent wants to act as the buyer’s agent as well, as it earns them double the commission, despite the obvious conflicts of interest. Of the agents we’d contacted previously, Masaru seemed the most trustworthy. I may have also been a bit biased, as he speaks English, whereas the other agents only spoke Japanese. He set up a viewing for us, but we discovered it wasn’t quite our ideal house. While the house was well built, the layout wasn’t our preference. It also was on a road with many large, noisy trucks, so it was kind of unpleasant to be outside. Making the first offer Over the next three months, we’d look at new listings online practically every day. I set up automated alerts for when the websites of different real estate agents were updated. We considered dozens of properties, and actually visited four plots and three used houses. None of them were a good match. On December 3rd, 2024, I found a great-looking property. It was being advertised as 5.1 km from Nakakaruiza station, which would have made it inconvenient, but it was also listed as 中部小学校近隣戸建 (Chubu Elementary Neighborhood Detached House). As the elementary school isn’t that far from the station, I assumed there was an error. The “house” was more of an old cabin, but it was priced cheaply for just the land itself. I sent Masaru an email asking him to give us the address. Meanwhile, my wife and I had gotten pretty good at identifying the locations of properties based on clues in the listing. The startup of a fellow Karuizawa international resident has a good tool for searching lots. My wife managed to identify this one, and so the following morning we looked at it. It wasn’t perfect, but it was a great location and a good deal, so before Masaru had even responded to us with the address, we sent another email saying we’d like to buy it. Masaru responded later that morning that he had a concern: the access to the house was a private road divided into two parcels, one owned by a real estate company, and the other by a pair of individuals. We’d need to negotiate with those parties about things like usage rights, waterworks, and so on. While he wasn’t worried about dealing with the real estate company, as it should just be a question of compensating them, the individuals owning the other parcel might theoretically be unwilling to negotiate. Based on his analysis, the following day (December 5th) we submitted a non-binding purchase application, with the condition that the road issue be resolved. That evening Masaru told us that he had discovered that the seller was a woman in her nineties. Because things like dementia are common among people that old, he added another condition, which was that a judicial scrivener testify to her mental capacity to enter into a contract. He warned that if we didn’t do that, and later the seller was to be found to be mentally unfit, the sale could be reversed, even if we had already proceeded with the construction of a new house. Given the cost of building a new home, this seemed like a prudent measure to take. A few days later we heard that the seller’s son was discussing the sale with his mother, and that they’d be having a family meeting the following weekend to decide whether to make the sale. To try to encourage them to sell to us, my wife wrote a letter (in Japanese) to them, and attached some photos of our family in Karuizawa. On December 16th, we received word that there was actually another potential buyer for the property who’d made an identical offer. In Japan, once someone submits a purchase application, it is normally handled on a first-come, first-served basis. However I noted that the seller didn’t mark the property as “under negotiation” on their website. I suspect we were the first to make an offer, as we did it literally the day after it was published. Apparently the seller’s son wanted to do business with us, but the daughter preferred the other family. We were not told why the daughter wanted to go with the other potential buyers, but it may have been because of our request to have the seller’s mental capacity tested. Dropping that requirement seemed too risky, though. On December 18th, we were informed that the seller had chosen the other buyer. This was quite a disappointing turn of events. We’d initially thought it was a sure thing because we were so fast to submit the offer and were giving them what they’d asked for. We’d even talked with a couple of home builders already. Furthermore, as winter had arrived, it seemed like the real estate market had also frozen. Whereas we’d previously seen new properties appear almost daily, now we spotted one or two a week at most. Making a successful offer Over the holidays, we looked at a couple more plots. On January 6th, we identified a used house in Nakakaruizawa that seemed to meet our criteria. It was by no means perfect, but close enough to be a candidate, so on January 10th we visited it. The house was pretty much as expected. Our one concern was the heating situation. It had previously had central heating that was based on a kerosene boiler, but that had been replaced with central air conditioning. While visiting the house, we turned it on, but it didn’t seem to make much difference. After some back and forth, we got the seller to run the AC constantly for a couple of days, and then on the morning of January 14th, we visited the house again. It was a grey and cold day, with little sunlight—basically the worst conditions possible for keeping the house warm, which I suppose was a good test. It was about 1°C outside and 10°C inside. We heard that the house had previously been rented out, and that the prior tenants had needed to use extra kerosene heaters to stay warm. This wasn’t ideal, but we also saw that with some renovations, we could improve the insulation. We considered making a lowball offer, but Masaru thought it was unlikely that it would be accepted. Instead, on January 17th, we made an offer at the asking price, conditional on it passing a home inspection and us getting an estimate for renovation costs. We heard that, once again, there was another potential buyer. They were “preparing” to submit an offer, but hadn’t done so yet. This time, the seller followed the standard process of considering our offer first, and accepting it in a non-binding manner. We learned that the seller itself was also a real estate company, whose main business seemed to be developing hotels and golf courses. The sale of this property was a relatively minor deal for them, and their main concern was that everything went smoothly. On January 23rd, we had a housing inspection carried out. The inspector found several issues, but they were all cosmetic or otherwise minor, so we said we’d proceed with the contract. Signing the contract The seller’s agent didn’t appear to be in a hurry to finalize the contract. It took a while for them to present us with the paperwork, and after that there were some minor changes to it, so it wasn’t until March 19th that we finally signed. One reason for the delay was that we’d suggested we sign the contract electronically, something the buyer had never done before. Signing a contract electronically was not only simpler from our perspective, but also avoided the need to affix a revenue stamp on the contract, saving some money. But even though Masaru was experienced with electronic contracts, it seemed to take the seller some time to confirm that everything was okay. When we signed the contract, we agreed to immediately make a 10 percent down payment, with the outstanding amount to be transferred by May 16th. We’d already gotten pre-approval on a housing loan, but needed the extra time for the final approval. Obtaining the loan On January 22nd, we visited Hachijuni Bank, which is headquartered in Nagano and has a branch in Karuizawa, and started the loan pre-approval process. This involved submitting a fair amount of paperwork, especially as I was the director of a company as opposed to a regular employee. The whole process took about two weeks. In parallel, we started the pre-approval process for PayPay Bank. The application involved minimal paperwork, and we were able to quickly get pre-approved by them too. To receive the actual approval for the loan, though, we needed to have already signed the purchase contract, so we couldn’t begin until March 19th. This time, the situation was reversed, and PayPay Bank required much more documentation than Hachijuni Bank. On April 9th, we met again with Hachijuni Bank. There was a problem with our application: my middle names hadn’t been properly entered into the contract, and so we’d have to go through the approval process again. The representative said he’d make sure they reviewed it quickly because it was his fault. That evening, we got the result of our application to PayPay Bank: rejected! No reason was given, but as they offer quite low rates, I’m guessing they only accept the lowest-risk loans. Maybe the reason was the house itself, maybe it was that I was the director of a small private company, maybe it was something else. While the Hachijuni Bank representative seemed positive about our application, I was worried. On April 15th we got the result from Hachijuni Bank: approved! Two days later we signed the paperwork with the bank and scheduled the handover for April 30th. Toilet troubles When the previous tenants had moved out of the house, the seller had decided to drain the water. Karuizawa temperatures are cold enough that if the house isn’t being heated, the pipes can freeze and burst. This meant that when we had the housing inspection performed, the inspector couldn’t confirm anything about the plumbing. Given that by mid-April temperatures were no longer below zero, we asked that the seller put the water back in. They agreed, and on April 23rd we heard that there was a problem with the toilets—some sort of leak. Likely some of the rubber sealing had dried out while the water had been removed. The seller offered to replace the toilets or pay us the cost of replacing them with equivalent ones. My understanding was that because they hadn’t declared any issues with the toilets in the original purchase agreement, they were obligated to do so. My wife and I had already talked about upgrading the toilets, but I wasn’t convinced that it’d be worth it. With the seller offering to pay two-thirds of the cost of the more premium toilets we wanted, though, it seemed like a stroke of luck. Handover On April 30th, the handover went smoothly. A representative from the seller’s company met with myself, the real estate agents, and a judicial scrivener. Technically I didn’t need to be there for it, but as the representative had come from Hiroshima I thought I should be present just in case anything happened. Nothing went awry, however, and they gave me the keys and some omiyage. Renovations We had already decided to do some renovations before moving in. With the toilets not working, that became an immediate necessity. One of the challenges of buying a used house is deciding how extensively to renovate. We settled on doing the minimum: replacing the toilets, putting up a wall to split the kids’ bedroom in two (it already had two separate doors and was designed to be split), switching the lights in the bedroom from fluorescent tubes to LEDs, and adding inner windows (内窓, uchi mado) to improve the insulation. We’d kicked off the process shortly after completing the home inspection, and had already had contractors visit to give estimates in advance. Still, we weren’t able to finalize any plans until we’d decided the exact handover date. After some negotiating, we got an agreement to wrap up the renovations by the beginning of June, and set a move-in date of June 5th. Moving in We managed to move in as scheduled, almost half a year after we first saw the house, and 10 months after we began our initial search. We weren’t exactly in a rush with the whole process, but it did take a lot of our time. Tips for potential homebuyers in Japan Here’s what I’d recommend if you’re looking to buy a house in Japan. Don’t rely on real estate agents to find properties for you My impression of real estate agents in Karuizawa is that they’re quite passive. I think this is a combination of the market being hot, and the fact that many inquiries they get are from people who dream about living here but will never make it a reality. I’d suggest being proactive about the search yourself. All of the best properties we found because we discovered the listings ourselves, not because an agent sent them to us. Find an agent you can trust The main value I got out of using a real estate agent was not because he introduced specific properties, but because he could point out potential issues with them. Checking the ownership of the road leading up to the property, or whether a contract could be reversed if the signer was years later deemed to be mentally incompetent, weren’t issues that had ever even occurred to me. Every property I considered purchasing had some unapparent problems. While there are some legal obligations placed upon the buyer’s agent, my understanding is that those don’t extend to every possible situation. What’s more, real estate seems to attract unscrupulous people, and some agents aren’t averse to behaving unethically or even illegally if it helps them close more deals. Because of this, I think the most important quality in an agent is trustworthiness. That can be hard to evaluate, but go with your gut. You don’t need your company to be profitable three years in a row I’d previously heard that it’s challenging to get a loan as the director of a company, and that the last three years of a business needed to be profitable to receive approval. That made me apprehensive, as TokyoDev was only incorporated in 2022, and 2024 was an unprofitable year for us due to it being the first year we had a decrease in the number of hires. However, when Hachijuni Bank reviewed TokyoDev’s finances, they said it was a great business. I was the sole employee of the business and it was only unprofitable because I had set my director’s compensation too aggressively. We’d had previous profitable years, and the company didn’t have any debts. In other words, one bad year for your company isn’t going to tank your chances of getting a loan. Go through the final approval process with multiple banks Some will advise you not to go through the final approval process of multiple banks at the same time. Since there’s a central database that allows banks to see that you’re doing this, it apparently increases your risk of rejection. If you’re only applying at a few banks, though, the risk remains fairly low. I guess PayPay Bank theoretically could have rejected us for this reason, but I doubt it. If we had initially applied only with them, however, and had waited to be rejected before applying with Hachijuni Bank, we would have risked not having the final approval go through before the payment deadline. Get a housing inspection If you’re buying a used house of any significant value, I highly recommend getting an inspection. We used Sakura Home Inspection, the biggest inspection company in Japan. While the base price for the inspection was only 66,000 yen, we ended up paying about 240,000 yen due to additional options, the size of the house, and transportation fees. While that’s not exactly cheap, it was less than one percent of the total purchase price, and seemed worth it to mitigate the risk of serious issues that might appear after the purchase. Rent locally before buying Our plan of first renting, then buying worked out quite well for us. Not only did it give me confidence that Karuizawa was somewhere I wanted to live for the long haul, it was much easier visiting potential properties when I lived a 10-minute bicycle ride away from them, instead of a 3-hour drive. If you’re considering buying a house in Karuizawa… Please get in touch with me! I’m happy to answer any questions you might have. In my previous article on coworking spaces in Karuizawa, I invited people to get in touch, and already have made several new connections through it. I’d love to continue to grow my community, and help transform Karuizawa from a resort town to one that’s focused on its full-time residents.
There’s a video on YouTube from “Technology Connections” — who I’ve never heard of or watched until now — called Algorithms are breaking how we think. I learned of this video from Gedeon Maheux of The Iconfactory fame. Speaking in the context of why they made Tapestry, he said the ideas in this video would be their manifesto. So I gave it a watch. Generally speaking, the video asks: Does anyone care to have a self-directed experience online, or with a computer more generally? I'm not sure how infrequently we’re actually deciding for ourselves these days [how we decide what we want to see, watch, and do on the internet] Ironically we spend more time than ever on computing devices, but less time than ever curating our own experiences with them. Which — again ironically — is the inverse of many things in our lives. Generally speaking, the more time we spend with something, the more we invest in making it our own — customizing it to our own idiosyncrasies. But how much time do you spend curating, customizing, and personalizing your digital experience? (If you’re reading this in an RSS reader, high five!) I’m not talking about “I liked that post, or saved that video, so the algorithm is personalizing things for me”. Do you know what to get yourself more of? Do you know where to find it? Do you even ask yourself these questions? “That sounds like too much work” you might say. And you’re right, it is work. As the guy in the video says: I'm one of those weirdos who think the most rewarding things in life take effort Me too. Email · Mastodon · Bluesky
