One of the interesting challenges of the AI ecosystem in 2026 is that new, effective patterns emerge faster than I can adopt them. I’ll find a handful, get back to work, and realize a month later that I’d missed four or five more. The adoption cycle for Imprint this year has been something like: January: get every engineer onto Claude Code every single day March: ok, let’s also get everyone else onto Claude Code or Claude Cowork every single day April: local development is bottlenecked on checkout and worktree model, instead create ~10 local workspaces which each have an independent checkout of every repository, and operate at the workspace level, not at the repository level, so it can generate cross-repository pull requests across frontend, backend, infrastructure and data monorepos June: oh boy, agent-driven development is heavily constrained by lack of a common task management system with higher visibility and less permission complexity than Jira, so let’s migrate the entire company over to Linear and hard stop on Jira July: yikes, now we have visibility into all these tickets, many of them are trivial but managing them through local development isn’t scaling, let’s roll out an orchestrated harness which internally we call “Agent Fleet”, along the lines of Stripe’s Minions The most recent question for me has been figuring out how to adopt the software factory pattern. (After some light research, the specific AI-context origin of this term is slightly messy to attribute, but I think it might be Justin McCarthy in February 2026’s Software Factories And The Agentic Moment.) The software factory pattern is looping on a broad goal, and then relying on the harness to drive progress towards that goal. Our first pass at implementation is fairly basic: An agent skill /linear-project-loop which reads in a Linear project and starts by auditing that project’s goal definition on these dimensions: An RFC in Notion that describes the project’s goals, how those goals are measured, and the general approach A Datadog dashboard or Snowflake queries that measure progress against those goals If those are missing, or the Linear project is missing in its entirety, it iterates with you on creating those missing tools. Then it reviews the state of the metrics and issues for the project. If new work is identified, it adds those issues to the project. It updates the state of issues that have moved. It works on the non-blocked tasks based on the project’s current state. This is often writing a pull request, updating a pull request, pinging for review, asking a clarifying question, etc. When a task completes, if the project description is fresh, it takes on the next task. If the description hasn’t been updated in a while, it reruns the loop starting with the first step. Right now I am running this locally in a local harness, but it’s working well enough that I anticipate moving the behavior to be driven by the same orchestrated harness that we assign one-off tasks to. What I particularly like about the factory pattern is that it parallels very closely how I’ve been working locally, while forcing me to recognize the places where I was accidentally hording parts of the state for myself regarding the goals of the project. I was already asking agents to iterate on specific Linear projects, but they didn’t have the ability to evaluate if they were going in the right direction, or if it was missing necessary tasks. Now it does. The other place this has been extremely helpful for me is checking in on projects post release. For example, I shipped our passkeys implementation earlier this year, but some months go by without my checking in on how it’s going. If we saw adoption spike, or error rates start to turn, I might miss it, but running the factory in a less frequent post-release mode would catch it immediately. The final thought that’s been interesting to me is how much all of the pieces here compound only to the extent that you have the other pieces. For example, this factory pattern depends on having Datadog MCP and Snowflake access available to manage goal-tracking, but it also depends on Linear being the single source of state for the company’s work, and an orchestrated harness that can perform work independently from your laptop. Keeping up with this many migrations is a fascinating industry moment.

Six years ago, I wrote Tech Lead Management roles are a trap. My argument then was that TLM roles present themselves as easier than moving into a full management role, but the tension between doing the software engineering and engineering management aspects of the role made being a TLM a much harder first management role than a pure engineering management role. I still agree with that post, and I have some additional bad news to share: middle management roles are mostly a trap as well, if you goal is to become an executive. The core aspects of middle management roles are: Balancing between top-down executive, lateral stakeholder, and bottom-up team pressure, e.g. keeping morale up as an organization deprioritizes last year’s big initiative Defining and operating an organization’s process, e.g. creating career ladders and interview loops Competing for a share of fixed organizational resources (e.g. budget) and allocating acquired resources These are all extremely important skills to be an effective executive, and they make up the bulk of The Engineering Executive’s Primer, but they are insufficient to make you a great executive. If you don’t have them, you will be a deeply flawed executive, but even if you’re an expert at them, you can still be a terrible executive. That’s because the most important skills of an effective executive are the same exact skills that make an excellent line manager: developing domain expertise, driving execution (including setting pace), and translating both of those into an organizational culture that extends beyond you (in any of innumerable different ways). All of them are more easily practiced and mastered as a line manager than as a middle manager. Most middle management roles make practicing those skills difficult, and sometimes negatively select against developing them. As a middle manager, if you drive execution too closely, you might get told off as a micromanager. As a middle manager, if you go too deep on domain expertise, you might get told that you’re not focusing enough on your internal stakeholders. That’s undoubtedly valid feedback in many middle management roles, but it’s the perfectly wrong feedback to someone who is trying to become an effective executive. As a result, I’ve come to believe that the filters for good middle managers inadvertently negatively select out the “challenging” line managers who actually have the best chance to be excellent executives. That’s not even necessarily irrational: most companies are developing middle managers to take on more complex middle management roles; very rarely do they worry about growing future executives from within. If you’re willing to embrace this fact that a little bit of time in middle management roles is important preparation to become an executive, but spending a great deal of time in middle management only prepares you for further middle management roles–and makes you less effective as a potential executive–then the AI-driven shift in managerial fads might be a threat to your current role, but it’s likely to improve your chances to succeed as an executive.

I’ve recently been thinking a lot about the concept of “soil horizons”, which is the idea that there are many distinct layers of soil, from topsoil all the way down to bedrock, which all combine into a soil horizon. Translating this idea into software, the ideal codebase would have a single uniform “code layer”, but a surprisingly large percentage of production software has numerous, distinct code layers as the leading architect shifted over time. I’ve found this particularly true for software in problem-spaces with high essential complexity and low scale complexity, where the purifying challenges of scaling never create enough pressure to compact disjoint layers into a unified layer. Codebases with the most code layers tend to be created by small teams working on complex domains over a long period of time. In many companies this might be an identity, permissions or payments team: stuff that’s permanently valuable, but usually not the central concern at any given time. On such teams, there is often only one architect who understands the nuances of the domain well enough to make tradeoffs. When that architect leaves, they are replaced by someone who aspires to operate in the same code layer, but simply cannot because they lack enough context to do so. As a result, that new replacement creates a new code layer, despite not intending to. If the team runs through a handful of folks as the new team leads struggle, it’s easy to end up with a complex code horizon very quickly. The problem of messy code horizons is not a new one, and the general approach to addressing them is the same one I wrote about seven years ago in Reclaim unreasonable software, but with the proliferation of coding and non-coding harnesses, lately I’m running into the problem of messy code horizons more frequently. Even more concerning, I’m seeing this problem expand from impacting code horizons into impacting how organizations make decisions outside of software, e.g. the company’s general reasoning horizons. When individuals or teams rely on LLMs to reason to conclusions, rather than using LLMs to explore or draft options, it’s possible for even the most important decisions to be built on top of flawed reasoning layers underneath. In the next section, I’ll develop the problem statement a bit about what I’m running into, and then in the final section I’ll lay out the approaches that I am finding (moderately) effective to navigate that problem. Messy reasoning horizons If you give three enthusiastic engineers a problem, a new codebase, a coding harness, and self-approval rights, it’s very easy to end up with three new soil horizons as their harnesses gleefully commit code. However, in engineering we have a number of techniques to derisk this problem. First, we have manual and automated code review, and second we increasingly have the ability for the harnesses to operate off sufficiently clear instructions that they write new code consistently with the existing code, even if the operator is unaware of what good looks like. This is also true for code review, where coding harnesses can drive consistency across pull requests even if the person (or harness) creating the pull requests is not operating off the same shared context as the wider team. Many codebases are not well-configured for this new reality, and those codebases are getting worse at an accelerating rate as more harness and agent contributions get added. Legacy codebases that reach a certain size before introducing these better practices are easier to fix than before, but still require a lot of work to fix. That said, I’m confident that coding harnesses are going to substantially improve the quality of code horizons over the next year or two as the way we configure harnesses improves. That’s not the problem I’m worried about. What I’m worried about is the application of harnesses to problems outside of writing software, where there’s no static typing, linting, or unit tests to validate the output. Let me provide a very recent example from my own work that highlights this problem: I wanted to understand how our incidents were trending over time. So I pulled data via an MCP, and the analysis was unintuitive to me, in particular I thought we were having more Data related incidents than the results reflected. I had to look at the incidents in Slack, then the results in our incident tool, and understand why the two conflicted. After a bit, I recognized the results in our incident tool were only showing incidents that properly tagged a team when the alert was triggered, so it was omitting about half the relevant incidents. After having the agent manually tag the incidents without team assignments, the data made a lot more sense. After recognizing the issue, it was trivial to fix. However, if I had simply accepted the initial analysis, I would have made the perfectly wrong conclusion about what was happening. On top of that wrong conclusion, I could have easily pushed the team to take on a project to solve an illusionary problem. What’s so pernicious about messy reasoning horizons, is once any reasoning layer is poisoned, it’s impossible to reason effectively on top of it. If you take the incident analysis example, it’s easy to imagine prioritizing the perfectly wrong set of remediations, which have the artifacts of solid strategic reasoning, but are nonetheless just wrong. It’s easy to imagine a team wasting a quarter of time building a solution to this sort of problem that never existed. It’s true that poor reasoning has always existed, long before harnesses, but my experience is that poor reasoning wearing well-formatted clothing is proliferating more widely than I’ve previously seen, and it is increasingly difficult to combat because certain social norms are – at least temporarily – collapsing around folks actually thinking. That collapse is largely driven by unprincipled adoption of AI techniques without paying attention to whether they work. Widespread adoption is, in my opinion, the fundamental risk for most companies at the moment, and something companies need to be doing, but many approaches inadvertently mix play (experimenting with something new in ways that are likely to fail!) with production (creating load-bearing work product!) in ways that erode social norms for quality. The norms are not uniformly collapsing by any means, they are generally intact, but even a small increase in the proliferation of low quality reasoning layers has a devastating effect on your ability to reason successfully. Especially true the further up the poor reasoning occurs (sloppy reasoning from senior leaders) or when senior leaders rely on layers of reasoning without inspection (leaders who aren’t sufficiently “in the details” to spot likely reasoning errors in reasoning layers). As a result, we now live in a world where accepting any part of the reasoning context before inspecting it might lead to making a catastrophic mistake. This is an exhausting way to live. Make no assumptions Accepting that this is the world we live in, I wanted to lay out the techniques that I am finding useful to deal with it. Some of these are novel, but many of them are the same techniques I was using before the LLM-advent: Make no assumptions. When new hires join my team or my company, the first thing I tell them is that it’s essential that they “make no assumptions.” This is difficult to do, and it goes against every instinct because it forces you to inspect each aspect of how the company works and thinks, but I do think it’s the necessary approach. It’s a bit like learning “internet-skepticism” at some point in your life, where you realize that everything on the internet is self-motivated in some way, and you have to maintain a strict filter on what ideas you accept. This is a hard change to make, but I genuinely believe this is the correct mindset for accepting new information in the current era. The combination of fewer management layers and more flawed reasoning layers means that the core job of leadership is inspecting the details. The author must be the first human in the loop for their output. The biggest cultural failure with harnesses is when you can tell that you–the recipient of a piece of work–are the first human in the loop reviewing it. You must set a cultural norm that the creator of a piece of content is always the first human in the loop before asking another human to review it. If you fail to set that cultural expectation, then you will quickly crush the remaining team with a high standard for quality reasoning, which will lead to a full destruction of your reasoning horizon. Prioritize reasonable software. Run the Reclaim unreasonable software playbook, recognizing that migrations are cheap in 2026, so it’s much faster to remediate gaps. The core idea here is that relying on convention doesn’t work, and instead you have to rely on deterministic decisioning for each approach. For humans this can feel overly prescriptive, but harnesses don’t care. Learn faster by separating play and production. Many folks trying to learn how to use harnesses and LLMs leap directly into using them in their most critical work. This is a slow way to learn, and can lead to substantial errors in your most critical work. It’s much faster to work by buffering small pockets of time to learn. For example, our head of data has spent time building an iOS app fully “hands off the keyboard” to get a better feel for the tools. This sort of experiment goes much faster and gives you more repetitions in less time. The very practical version of this is setting aside a day or two periodically for folks to experiment. Structure how you think with LLMs. In Crafting Engineering Strategy, I lay out a structured approach to reasoning through creating a strategy document, which aims to prevent the reasoning errors that folks make in their thinking. This applies equally in how we use LLMs, and I think you can substantially reduce the chance of introducing flawed reasoning layers by focusing LLM work on exploration (gathering information on internet and via various MCPs), refinement (presenting gathered information effectively), and a final formatting pass. That takes much of the work out of strategy creation while constraining the areas you have to avoid making any assumptions about its output. I’m certain there are more things! What are you trying?

From early 2014 through late 2020, I was working in hypergrowth environments, which are challenging, but also educational. The most valuable feature of hypergrowth is that your mistakes reveal themselves next month rather than next year, because things go wrong very loudly when you’re moving fast. I’ve been thinking a lot about hypergrowth recently, because Imprint’s business is growing quickly and we did a large batch of hiring last year, but also because the AI-tooling shift has changed the pace at which it’s possible to work. This post documents the new rules I’ve revised my approach to engineering leadership around, and then talks through the specific projects I’ve worked on over the past year that caused me to believe in these rules. Revised rules Migrations can be done by an individual rather than a team. Even complex, large changes can be 95% owned by the driving individual or team, and done in 10% of the time. As the initial cost of migrations goes down, the reward/penalty of each migration’s quality goes up: even small sharp edges will break your colleagues’ mental models about the software you co-maintain. The impact of individual judgment on your company has never been higher. While 1st-pass code is nearly free, the cost of working code depends on your development harness, and is not free. We’re in an era when many companies say that everyone should be writing code, however our experience is that writing code that works well, while avoiding messy edgecases, remains difficult. Just how difficult remains a factor of your development harness, e.g. your tests, CI/CD, validation environments, preview-ability of changes, and so on. While I personally don’t imagine it’s valuable for most folks at a company to be contributing code, I suspect that most disagreement about that topic is actually a miscommunication: even at a company where “everyone codes”, the marketing team isn’t reducing allocations in your servers, instead it’s about whether there is a safe boundary where they can participate. (Much like a SaaS product that allows customization by writing software.) The good news is that this means the things that were most valuable to speed up engineering two years ago are still the things that are most valuable to speed them up today. Optimize the base-case of process for agents. Most steps of most processes can be fully automated in most cases. With the right harnesses, the right controls, domain context, and good judgment in their designers, you can fully automate the base-case of most processes in modern technology companies. For example, the base case of code review from a human is slower and less effective than a good harness’ code review. Of course, the harness will miss things, but so will human reviewers, and most areas are relatively safe to make changes. Of course, there are some higher risk areas, where this doesn’t hold true. By effectively capturing these distinctions properly, we can go much faster without introducing risk. By failing to capture these distinctions, we’ll create innumerable problems for ourselves. As a corollary, I think most planning processes like weekly or bi-weekly sprints are operating at too low an altitude. Humans planning together still matters, but should be operating at a higher level. Durable, high-ownership teams with domain-context are even more important. One of my biggest lessons at Uber was that persistent, durable teams work magic by accumulating domain-context, building a sense of camaraderie, and feeling an increasingly strong sense of ownership over an area as they continue to work in it. Even in an era where specifically doing something is much cheaper, you still have to do the right thing, which has gotten a bit easier but not much easier, and structural improvements help address this. (As a recent example of that, we had an issue in production where the necessary data to optimize it simply wasn’t being captured at all, so the harness’ ideas to solve it were reasonable but wrong, since the only real path forward was instrumenting the missing information.) As a specific disagreement, there’s a prevailing idea that AI-first companies will be run by a small number of genius engineers who create perfect versions of things one by one, doing such a good job that there’s nothing to maintain. This is a very compelling vision, but I don’t see it happening. High judgment individuals can wander across a company doing remarkable things, but at some point they do get hemmed in by lack of domain context, which is why durable teams are the fundamental building block, even in this era. Quick, good, and durable decision-making is a prerequisite to meaningfully benefit from AI. Being able to replace a legal review with automation only works if Legal can commit to that change, which depends on designing the automation thoughtfully, and also the teams’ willingness to collaborate. Implementing a new feature is only valuable if you can decide to launch that feature. Your team and company can only benefit from this increased pace of execution if you can make durable decisions quickly, and those decisions are good. This is the primary reason, in my opinion, why the average CTO role has necessarily become substantially more technical and less bureaucratic than a year ago. In many cases, I am the only person who can make binding decisions when teams disagree on the path forward, and that means I am making decisions constantly in this new world in order to maintain the pace. (That’s not an argument that executives are better decision makers, just that binding executive decisions are uniquely powerful to the extent that the executives themselves are aligned enough to honor those decisions.) What have we done in practice So, I genuinely believe the above rules based on my experiences over the past year, and let me try to connect them to specific projects we’ve worked on that have convinced me of them: Migrations A year ago, we deployed manually, and deployed ~6 times a week, and now we deploy 200-400 times a week. Our engineering headcount has doubled, but even if we double the prior deploys, we’re still up 20-30x year-over-year. This is due to a complete overhaul of how we deploy and run migrations, and this migration was done over two months and done 90% by two folks on our infrastructure team. The first day of January, about 25% of folks on our team used Claude Code or Cursor every day. By the end of February, 100% did. We did this without any top-down mandate, just by making the tooling good and chatting with non-adopters to remove sources of friction. Pretty much every PR is written by harnesses now, at least in the first pass. We migrated from a large number of varied configuration mechanisms to two configuration mechanisms (one for client or server constants that rarely change, a second for product-specific or frequently changing values). This was a large series of changes, which were largely done as a series of isolated projects by individual engineers. First, one engineer cleaned up the architecture to support this approach. Then another engineer did a reference architecture on the new approach. Then several more engineers followed the reference architecture in other areas of our codebase. This might have been a years long project of many people in the prior world, but took less than a quarter to complete, including a new internal tool for managing these values across engineering and non-engineer teams. We unified a multi-repo frontend application architecture into a mono-repo frontend architecture over about a month. This was 95% driven by one frontend engineer. We now have a shared frontend development harness, can maintain libraries cheaply, and entirely moved off using npm for package hosting, which was a source of ongoing friction. We fully statically typed our frontend code, going from a place where the majority of our frontend code was not typed. This was done by one engineer, and a lot of tokens, over the course of a few weeks. We migrated from npm to pnpm for better security defaults and faster deploys. This took one engineer a few hours a day for a few days. Cost of working code depends on your development harness. Where we’ve tried to throw design documents and PRs “over the wall” to engineers on other teams, they’ve never gone anywhere. Slop pull requests and design documents are cheap, but are actively harmful. They not only have to be cleaned up and repaired, their context poisons the LLM, leading to worse outcomes than starting over. We’ve seen tremendous success in managers contributing software, as long as those managers are validating the work directly, looking at dashboards after their changes go out, and resolving any issues their changes cause. We’ve found no positive impact from folks attempting to make changes where they don’t do those things. Optimize the base-case of process for agents. We triage all incoming issues from our customer operations team using a harness which knows our team, our open tickets, and has limited access to our data warehouse to size the impact of issues. This is complex, high-skill but not particularly interesting labor that we’re now doing better and faster with agents. Yes, there is still a human triage for the edgecases. Importantly, we’re also doing this without changing human workflows, it’s the same workflow, just with some steps automated. The first pass of code review is done by the same harness that implements the changes, cleared of the context used to write the change, allowing humans to focus on higher value feedback. We rolled out Claude Code and Cowork to all folks in the company last quarter, and have seen them also automate an increasingly large swath of their work as well. Our fraud team has been particularly ambitious in replacing manual workflows with a first-pass of automation–with attribution to the data itself–to do the initial investigation on potential attacks automatically. We’ve migrated to Linear, and off Jira, to better support this workflow with a more capable MCP and better Slack integration, making it possible for everyone internally to have better infrastructure for building these agent-first workflows. More on this later, but we’re almost done alpha-testing our internal harness pulling issues off Linear, and working to resolve them, automatically which is our biggest next step in this direction. Durable, high ownership teams with domain-context are even more important. When I joined, we had a number of areas supported by very talented folks who rotated through them quickly on a per-project basis. This worked, but it meant we were very reactive to issues. Now, we’ve been able to dedicate at least a small team to every important area of the company, where they are able to persistently invest. These teams are now wielding all the new techniques afforded by AI themselves. Without them, no one would be capturing these opportunities, because there is simply too much happening. We launched SierraAI, which is quite good, but since then the team has iterated on it relentlessly, getting it truly excellent. This is something we wouldn’t have been able to do without a dedicated, focused team. Quick, good and durable decision making is a prerequisite to benefit from AI. Changing how we do configuration was a controversial decision, and I’ve had to make repeated clarifications on the approach. This would have been very difficult to do bottom-up, because it impacts every team differently, and the benefit is only experienced at the ecosystem-level (allowing one person to configure all configuration across teams). Reworking our CI/CD pipeline was controversial, as it changed many folks’ mental models of how we deploy and release (e.g., it forced us to explicitly decouple deploy and release via feature flagging). This was a contentious decision, and would have been slow and difficult to make bottom-up. Unifying into a web mono-repo was also a controversial decision with varied opinions. It benefitted greatly from having a unified decision. Moving to SierraAI was a difficult discussion versus both various competitors, and also not doing it. It needed the executive stamp to finalize the cross-functional debate. These are just representative examples, we’ve done a lot more than these. The aperture of what’s possible has continued to expand every month this year, but the things holding us back haven’t changed all that much: organizational misalignment, lack of clarity, and poor technical architecture. It’s a wild time to be working in technology.

One of my gifts/curses is an endless fixation with how processes can be optimized. For a brief moment early in my career, that was focused on improving how humans collaborate, but that quickly switched to figuring out how we can minimize human involvement, and eliminate human-to-human handoffs as much as possible. Lately, every time I perform a recurring task–or see someone else perform one–I think about how we might eliminate the human’s involvement entirely by introducing agents. This both has worked well, but also worked poorly, and I wanted to highlight the pattern I’ve found useful. For a concrete example, a problem that all software companies have is patching security vulnerabilities. We have that problem too, and I check our security dashboards periodically to ensure nothing has gone awry. Sometimes when I check that dashboard, I’ll notice a finding that’s precariously close to our resolution SLAs, and either fix it myself or track down the appropriate team to fix it. However, this feels like a process that shouldn’t require me checking on it. Five to six months ago, I added Github Dependabot webhooks as an input into our internal agent framework. Then I set up an agent to handle those webhooks, including filtering incoming messages down to the highest priority issues. About a month ago, when I upgraded from GPT 4.1 to GPT 5.4 with high reasoning, I noticed that it got quite good at using the Github MCP to determine the appropriate owners for a given issue, using the same variety of techniques that a human would use: looking at Codeowners files where available, looking at recent commits on the repository, and so on. The alerts and owners were already getting piped into a Slack channel. So, this worked! However, it didn’t actually work that well, because despite repeated iteration on the prompt, including numerous CRITICAL: you must... statements, it simply could not reliably restrict itself to critical severity alerts. It would also include some high severity alerts, and even the occasional medium severity alert. This is a recurring issue with using agents as drop-in software replacement: they simply are not perfect, and interrupting your colleagues requires a level of near-perfection. If I’d hired someone on our Security team to notify teams about critical alerts, and they occasionally flagged non-critical alerts, eventually someone would pop into my DMs to ask me what was going wrong. That didn’t happen here, because the knowledge that those DMs would show up prevented me from rolling the notifications out more aggressively. Coding agents address this sort of issue by running tests, typechecking, or linting, but less structured tasks are either harder or more expensive to verify. For example, I could have added an eval verifying messages didn’t mention medium or high severity tasks before allowing it to send to Slack, but I found that somewhat unsatisfying despite knowing that it would work. Instead, after some procrastination on other tasks, I finally prompted Claude to update this agent to rely on a code-driven workflow where flow-control is managed by software by default, and only cedes control to an agent where ideal. That workflow looks like: A webhook comes in from Dependabot Script extracts the severity and action (e.g. is it a new issue versus a resolved issue), and filters out low priority or non-actionable webhooks The code packages the metadata into a list of issues and repositories The code passes each repository-scoped bundle to an agent with our internal ownership skill and the Github MCP to determine appropriate folks to notify for each issue The issues and ownership data are passed to a second agent that formats them as a Slack message This works 100% of the time, while still allowing us to rely on our internal ownership skill to determine the most likely teams or individuals to notify for a given problem. It’s now something I can rollout more aggressively. The immediate fast follow was a weekly follow-up ping for open critical issues, relying on the same split of deterministic and agentic behaviors. The next improvement will be automating the generation of the vulnerability fixes, such that the human involvement is just reviewing the change before it automatically deploys. (We already do this for Dependabot generated PRs, but in my experience Dependabot can solve a reasonable subset of identified issues, but far from all of them.) That is the pattern that I’ve found effective: Prototype with agent-driven workflow until I get a feel for the workflow and what’s difficult about it Refactor agent-driven control away, increasingly relying on code-driven workflow for more and more of the solution End with a version that narrowly relies on agents for their strengths (navigating ambiguous problems like identifying code owners) This has worked well for pretty much every problem I’ve encountered. The end-result is faster, cheaper, and more maintainable. It’s also a cheap transition, generally I can take logs of some recent runs, the agent’s prompt, and some brief instructions, throw them into Codex/Claude, and get a working replacement in a few minutes.