It's true that in my company we're not building rockets or defense systems, maybe you guys are and in those scenarios it's less useful. But for typical LoB and/or consumer-facing software, AI is crushing it. Where I used to need 3 devs, now I just need one (and the support team around it: PM, BA, QA, Designer). For my business, AI has been a game changer.
agents role (Orchestrator, QA etc.), agents communication, thinking patterns, iteration patterns, feature folders, time-aware changelog tracking, prompt enforcing, real time steering.
We might really need a public Wiki for that (C2 [1] style)
[1]: https://wiki.c2.com/
- Through the last two decades of the 20th century, Moore’s Law held and ensured that more transistors could be packed into next year’s chips that could run at faster and faster clock speeds. Software floated on a rising tide of hardware performance so writing fast code wasn’t always worth the effort.
- Power consumption doesn’t vary with transistor density but varies with the cube of clock frequency, so by the early 2000s Intel hit a wall and couldn’t push the clock above ~4GHz with normal heat dissipation methods. Multi-core processors were the only way to keep the performance increasing year after year.
- Up to this point the CPU could squeeze out performance increases by parallelizing sequential code through clever scheduling tricks (and compilers could provide an assist by unrolling loops) but with multiple cores software developers could no longer pretend that concurrent programming was only something that academics and HPC clusters cared about.
CS curricula are mostly still stuck in the early 2000s, or at least it feels that way. We teach big-O and use it to show that mergesort or quicksort will beat the pants off of bubble sort, but topics like Amdahl’s Law are buried in an upper-level elective when in fact it is much more directly relevant to the performance of real code, on real present-day workloads, than a typical big-O analysis.
In any case, I used all this as justification for teaching bitonic sort to 2nd and 3rd year undergrads.
My point here is that Simon’s assertion that “code is cheap” feels a lot like the kind of paradigm shift that comes from realizing that in a world with easily accessible massively parallel compute hardware, the things that matter for writing performant software have completely shifted: minimizing branching and data dependencies produces code that looks profoundly different than what most developers are used to. e.g. running 5 linear passes over a column might actually be faster than a single merged pass if those 5 passes touch different memory and the merged pass has to wait to shuffle all that data in and out of the cache because it doesn’t fit.
What all this means for the software development process I can’t say, but the payoff will be tremendous (10-100x, just like with properly parallelized code) for those who can see the new paradigm first and exploit it.
And actually, these tools actually work, , because 99% of people still don’t really know how to prompt agents well and end up doing things like “pls fix this, it’s not working”.
One thing that worked well for us was going back to how a human team would approach it: write a product spec first (expected behavior, constraints, acceptance criteria, etc), use AI to refine that spec, and only then hand it to an opinionated flow of agents that reflect a human team to implement.
Not hypothetically -- like, actually does it. Ships the quarterly analysis, writes the strategy memo, builds the model. All in a fraction of the time. And then management finds out.
There's a short film that nails this exact scenario and it's weirdly accurate for anyone who's worked in a large org: https://youtu.be/O5FFkHUdKyE
The engineering patterns will keep evolving. But the harder problem is what organizations do when one person with an AI agent outperforms a department. That's not a technical problem -- it's a people problem.
So far I only have one: Inflicting unreviewed code on collaborators, aka dumping a thousand line PR without even making sure it works first https://simonwillison.net/guides/agentic-engineering-pattern...
Test fail -> implement -> linter -> test pass
Another idea I've thought about using is docs driven development. So the instructions might look like..
Write doc for feat/bug > test fail > implement > lint > test pass
Other things that I feel are useful:
- Very strict typing/static analysis
- Denying tool usage with a hook telling the agent why+what they should do (instead of simple denial, or dangerously accepting everything)
- Using different models for code review
This brings the Linux Kernel style patch => discuss => merge by maintainer workflow to agents. You get bisect safe patches you 'review' and provide feedback and approve.
While a SKILL could mimic this, being built in allows me to place access control and 'gate' destructive actions so the LLM is forced to follow this workflow. Overall, this works really well for me. I am able to get bisect-safe patches, and then review / re-roll them until I get exactly what I want, then I merge them.
Sure this may be the path to software factories, but it scales 'enough' for medium size projects and I've been able to build in a way that I maintain strong understanding of the code that goes in.
For a high level description of what this new way of engineering is about: https://substack.com/@shreddd/p-189554031
The damn thing _talks_. You can just _speak_ to it. You can just ask it to do what you want.
Teams jump straight to multi-agent setups because it feels "advanced," then spend weeks debugging ghosts. A boring single-agent flow with strict tool contracts, replayable traces, and a small regression suite beats a clever architecture you can't reason about.
My rule of thumb now: if I can't explain why an agent made a decision from logs alone, it's not production-ready yet.
In one of my experiments I had the simple goal of "making Linux binaries smaller to download using better compression" [1]. Compression is perfect for this. Easily validated (binary -> compress -> decompress -> binary) so each iteration should make a dent otherwise the attempt is thrown out.
Lessons I learned from my attempts:
- Do not micro-manage. AI is probably good at coming up with ideas and does not need your input too much
- Test harness is everything, if you don't have a way of validating the work, the loop will go stray
- Let the iterations experiment. Let AI explore ideas and break things in its experiment. The iteration might take longer but those experiments are valuable for the next iteration
- Keep some .md files as scratch pad in between sessions so each iteration in the loop can learn from previous experiments and attempts
One of the biggest struggles I have on my team is coworkers straight up vibing parts of the code and not understanding or guiding the architecture of subsystems. Or at least, not writing code in a way that is meant to be understood by others.
Then when I go through the code and provide extensive feedback (mostly architectural and highlighting odd inconsistencies with the code additions) I'm met with much pushback because "it works, why change it"? Not to mention the sheer size of prs ballooning in recent months.
The end result is me being the bottleneck because I can't keep up with the "pace" of code being generated, and feeling a lot of discomfort and pressure to lower my standards.
I've thought about using a code review agent to review and act as me in proxy, but not being able to control the exact output worries me. And I don't like the lack of human touch it provides. Maybe someone has advice on a humane way to handle this problem.
Shameless plug: I wrote one. https://marmelab.com/blog/2026/01/21/agent-experience.html
"deeply understand this codebase, clearly noting async/sync nature, entry points and external integration. Once understood prepare for follow up questions from me in a rapid fire pattern, your goal is to keep responses concise and always cite code snippets to ensure responses are factual and not hallucinated. With every response ask me if this particular piece of knowledge should be persistent into codebase.md"
Both the concise and structure nature (code snippets) help me gain knowledge of the entire codebase - as I progressively ask complex questions on the codebase.
As my projects were growing in complexity and scope, I found myself worrying that we were building things that would subtly break other parts of the application. Because of the limited context windows, it was clear that after a certain size, Claude kind of stops understanding how the work you're doing interacts with the rest of the system. Tests help protect against that.
Red/green TDD specifically ensures that the current work is quite focused on the thing that you're actually trying to accomplish, in that you can observe a concrete change in behaviour as a result of the change, with the added benefit of growing the test suite over time.
It's also easier than ever to create comprehensive integration test suites - my most valuable tests are tests that test entire user facing workflows with only UI elements, using a real backend.
https://simonwillison.net/guides/agentic-engineering-pattern...
I am still not sold on agentic coding. We’ll probably get there within the next couple of years.
Colleagues don’t usually like to review AI generated code. If they use AI to review code, then that misses the point of doing the review. If they do the review manually (the old way) it becomes a bottleneck (we are faster at producing code now than we are at reviewing it)
- tell the agent to write a plan, review the plan, tell the agent to implement the plan
- allow the agent to “self discover” the test harness (eg. “Validate this c compiler against gcc”)
- queue a bunch of tasks with // todo … and yolo “fix all the todo tasks”
- validate against a known output (“translate this to rust and ensure it emits the same byte or byte output as you go”)
- pick a suitable language for the task (“go is best for this task because I tried several languages and it did the best for this domain in go”)
> A comprehensive test suite is by far the most effective way to keep those features working.
IMO this section is a great place to show how we as humans can guide the LLM toward a rigorous test suite, rather than one that has a lot of "coverage" but doesn't actually provide sound guarantees about behavior.
Feels like it’s a lot of words to say what amounts to make the agent do the steps we know works well for building software.
The thing I keep coming back to is that it's all code. Almost all white collar professions have at least some key outputs in code. Whether you are a store manager filling out reports or a marketing firm or a teacher, there is so much code.
This means you can give claude code a branded document template, fill it out, include images etc. and uploaded to our cloud hosting.
With this same guidance and taste, I'm doing close to the work of 5 people.
Setup: Claude code with full API access to all my digital spaces + tmux running 3-5 tasks in parallel
I've dipped into agentic work now and again, but never been very impressed with the output (well, that there is any functioning output is insanely impressive, but it isn't code I want to be on the hook for complaining).
I hear a lot of people saying the same, but similarly a bunch of people I respect saying they barely write code anymore. It feels a little tricky to square these up sometimes.
Anyway, really looking forward to trying some if these patterns as the book develops to see if that makes a difference. Understanding how other peopke really use these tools is a big gap for me.
Like an engineer overseeing the construction of a bridge, the job is not to lay bricks. It is to ensure the structure does not collapse.
The marginal cost of code is collapsing. That single fact changes everything.
like don't ask it to "write tests for this function", instead give it a function that's deliberately broken in a specific way, make it write a test that catches that bug, verify the test actually fails, THEN fix the function
this forces the test to be meaningful because it has to detect a real failure mode. if the agent can't make the test fail by breaking the code, the test is useless
the other thing that helps is being really specific about edge cases upfront. instead of "write tests for this API endpoint", say "write tests that verify it returns 400 when the email field is missing, returns 409 when the email already exists, returns 422 when the email is malformed" etc
agents are weirdly good at implementing specific test scenarios but terrible at figuring out what scenarios actually matter. which honestly is the same problem junior devs have lol
Take a guitar, for example. You don't industrialize the manufacture of guitars by speeding up the same practices that artisans used to build them. You don't create machines that resemble individual artisans in their previous roles (like everyone seems to be trying to do with AI and software). You become Leo Fender, and you design a new kind of guitar that is made to be manufactured at another level of scale magnitude. You need to be Leo Fender though (not a talented guitarrist, but definitely a technical master).
To me, it sounds too early to describe patterns, since we haven't met the Ford/Fender/etc equivalent of this yet. I do appreciate the attempt though.
I distilled multiple software books into these flows and skills. With more books to come.
Here is an example https://github.com/ryanthedev/code-foundations
> You can just ask it to do what you want.
Yes, but very clearly, as any HN thread on AI shows, different people are having VERY different outcomes with it. And I suspect it is largely the misconception that it will magically "just do what you want" that leads to poor outcomes.
The techniques mentioned -- coding, docs, modularity etc. -- may seem obvious now, but only recently did we realize that the primary principle emerging is "what's good for humans is good for agents." That was not at all obvious when we started off. It is doubly counter-intuitive given the foremost caveat has been "Don't anthropomorphize AI." I'm finding that is actually a decent way to understand these models. They are unnervingly like us, yet not like us.
All that to say, AI is essentially black magic and it is not yet obvious how to use it well for all people and all use-cases, so yes, more exposition is warranted.
I think attempts to document the most effective things to ask it to do in order to get to your overall goal, as well as what it is and is not good for, is probably worth doing. It would be bad if it turned into a whole consultant marketing OOP coaching clusterfuck. Yeah, but building some kind of community knowledge that these things aren't like, demigods, they have limitations and during things one way or the other with them can be better is probably a good thing. At the very least in theory would cut down some of the hype?
It's not as simple an observation as you're making it out to be.
I suspect that this time around, management will expect the AI chatbot to explain these things to you, because who pays for anything anymore if the AI can do it all.
Good news - agents are good at open ended adding new tests and finding bugs. Do that. Also do unit tests and playwright. Testing everything via web driving seems insane pre agents but now its more than doable.
This is the most important piece to using AI coding agents. They are truly magical machines that can make easy work of a large number of development, general purpose computing, and data collection tasks, but without deterministic and executable checks and tests, you can't guarantee anything from one iteration of the loop to the next.
A broken test doesn’t make the agentic coding tool go “ooooh I made a bad assumption” any more than a type error or linter does
All a broken test does it prompt me to prompt back “fix tests”
I have no clue which one broke or why or what was missed, and it doesnt matter. Actual regressions are different and not dependent on these tests, and I follow along from type errors and LLM observability
There's a lifecycle to these hype runs, even when the thing behind the hype is plenty real. We're still in the phase where if you criticize AI you get told you don't "get it", so people are holding back some of their criticisms because they won't be received well. In this case, I'm not talking about the criticisms of the people standing back and taking shots at the tech, I'm talking about the criticisms of those heavily using it.
At some point, the dam will break, and it will become acceptable, if not fashionable, to talk about the real problems the tech is creating. Right now there is only the tiniest trickle from the folk who just don't care how they are perceived, but once it becomes acceptable it'll be a flood.
And there are going to be problems that come from using vast quantities of AI on a code base, especially of the form "created so much code my AI couldn't handle it anymore and neither could any of the humans involved". There's going to need to be a discussion on techniques on how to handle this. There's going to be characteristic problems and solutions.
The thing that really makes this hard to track though is the tech itself is moving faster than this cycle does. But if the exponential curve turns into a sigmoid curve, we're going to start hearing about these problems. If we just get a few more incremental improvements on what we have now, there absolutely are going to be patterns as to how to use AI and some very strong anti-patterns that we'll discover, and there will be consultants, and little companies that will specialize in fixing the problems, and people who propose buzzword solutions and give lots of talks about it and attract an annoying following online, and all that jazz. Unless AI proceeds to the point that it can completely replace a senior engineer from top to bottom, this is inevitable.
COBOL's promise was that it was human-like text, so we wouldn't need programmers anymore.
The problem is that the average person doesn't know how what their actual problems are in sufficient detail to get a working solution. When you get down to breaking down that problem... you become a programmer.
The main lesson of COBOL is that it isn't the computer interface/language that necessitates a programmer.
yes. It sucks but I think it's good for the next generation of tech industry employees to watch this. It's happening quickly so you get a 10 year timeline compressed into a few years which makes it easier to follow and expose. The bloggers will come, then speakers, then there will be books. Consultants will latch on and start initiatives at their clients. Once enough large enterprises are sold on it, there will come associations and certification bodies so a company can say "we have X certified abc on staff". Manifestos will be released, version numbers will be incremented so there's a steady flow of work writing books, doing trainings, and getting the next level certified.
This is standard issue tech industry stuff (and it probably happens everywhere else too) but compressed into a tighter timeline so you don't have to wait a decade to see it unfold.
The context suggests the former, but your criticisms bear no relation to the linked content. If anything, your edict to "write tests first" is even more succinctly expressed as "Red/green TDD".
The tricky part in our case is that "behaves correctly" has two layers - functional (did it navigate correctly?) and behavioral (does it look human to detection systems?). Agents are fine with the first layer but have no intuition for the second. Injecting behavioral validation into the loop was the thing that actually made it useful.
The .md scratch pad between sessions is underrated. We ended up formalizing it into a short decisions log - not a summary of what happened, just the non-obvious choices and why. The difference between "we tried X" and "we tried X, it failed because Y, so we use Z instead" is huge for the next session.
summaries ("tried X, tried Y, settled on Z") are better than nothing, but the next iteration can mostly reconstruct them from test results anyway. what's actually irreplaceable is the constraint log: "approach B rejected because latency spikes above N ms on target hardware" means the agent doesn't re-propose B the next session. without it, every iteration rediscovers the same dead ends.
ended up splitting it into decisions.md and rejections.md. counter-intuitively, rejections.md turned out to be the more useful file. the decisions are visible in the code. the rejections are invisible — and invisible constraints are exactly what agents repeatedly violate.
0: https://wiki.roshangeorge.dev/w/Blog/2025-12-01/Grounding_Yo...
Which is oddly close to how investment advice is given. If these techniques work so well, why give them up for free?
A lot of Simon's tools are making harnesses for this so it can get integrated:
showboat - create a demo, validate the code generates the demo. This is making a documentation source of truth
rodney - validate visually and with navigation that things work like you expect
red-green tests are conceptually the same - once we have these tests then the agent can loop more successfully.
So, I think there are some "universals" or at least "universals for now" that do transcend team/deployment specificity
The problem I kept hitting was that flat markdown constraint logs don't scale past ~50 entries. The agent has to re-read the entire log to know what was already tried, which eats context window and slows generation. And once you have multiple agents in parallel, each maintaining their own constraint log, you get drift - agent A rejects approach B, agent C re-proposes it because it never saw agent A's log.
What worked for me was moving constraint logs to append-only log blocks that agents query through MCP rather than re-read as prose. I've been using ctlsurf for this - the agent appends 'approach B rejected, latency > N ms' to a log block, and any agent can query query_log(action='approach_rejected') to see what's been ruled out. State store handles 'which modules are claimed' as a key-value lookup.
Structured queries mean agents don't re-read the whole history - they ask specific questions about what's been tried.
Basically, it's Waterfall for Agents. Lots of Capitalized Words to signify something.
Also they constantly call it the BMAD Method, even though the M already stands for method.
- A lot more linting rules than ever before, also custom rule sets that do more org and project level validations.
- Harder types enforcement in type optional languages , Stronger and deeper typing in all of them .
- beyond unit tests - test quality coverage tooling like mutation testing(stryker) and property based testing (quickcheck) if you can go that precise
- much more dx scripts and build harnesses that are specific to org and repo practices that usually junior/new devs learn over time
- On dynamic side , per pull requests environments with e2e tests that agents can validate against and iterate when things don’t work.
- documentation generation and skill curation. After doing a batch of pull requests reviews I will spend time in seeing where the gaps are in repo skills and agents.
All this becomes pre-commit heavy, and laptops cannot keep up in monorepos, so we ended up doing more remote containers on beefy machines and investing and also task caching (nx/turborepo have this )
Reviews (agentic or human) have their uses , but doing this with reviews is just high latency, inefficient and also tends to miss things and we become the bottleneck.
Earlier the coder(human or agent) gets repeatable consistent feedback it is better
Also, we only allow engineers to commit (agent generated) code. Designers just come up with suggestions, engineers take it and ensure it fits our architecture.
We do have a huge codebase. We are teaching Claude Code with CLAUDE.md's and now also <feature>.spec.md (often a summary of the implementation plan).
In the end, engineers are responsible.
This is an educational problem, and is unlikely to be easy to fix in your team (though I might be wrong). I would suggest to change to a team or company with a culture that values being able to reason about one’s software.
There are plenty of orgs using AI who still care about architecture and having easily human-readable, human-maintainable code. Maybe that's becoming an anachronism, and those firms will go the way of the Brontosaurus. Maybe it will be a competitive advantage. Who knows?
¹ "Make it work, make it right, make it fast."
But more proactively, if people aren't going to write their own code, I think there needs to be a review process around their prompts, before they generate any code at all. Make this a formal process, generate the task list you're going to feed to your LLM, write a spec, and that should be reviewed. This is not a substitute for code reviews, but it tends to ensure that there are only nitpick issues left, not major violations of how the system is intended to be architected.
>but not being able to control the exact output worries me
Why?
https://agentexperience.ax/ describes it as "refers to the holistic experience AI agents have when interacting with a product, platform, or system" which feels to me like a different concept to figuring out patterns for effectively using coding agents as a software engineer.
You can "pip install ziglang" and get the right version for different platforms too.
Worse: once you have one "bad apple" in your pile of tests, it decreases trust in the _whole batch of tests_. Each time a test passes, you have to think if it's a bad test...
I have a hard enough time getting humans to write tests like this…
"Explain the codebase to a newcomer. What is the general structure, what are the important things to know, and what are some pointers for things to learn next?"
Once I saw the output I giddyup'd and haven't looked back.
Dismissing everything AI as slop strikes me as an attitude that is not going to age well. You’ll miss the boat when it does come (and I believe it already has).
It's tricky though. Take "red/green TDD" for example - it's perfectly possible that models will start defaulting to doing that anyway pretty soon.
In that case it's only three words so it doesn't feel hugely wasteful if it turns out not to be necessary - and there's still value in understanding what it means even if you no longer have to explicitly tell the agents to do it.
Claude Code experience is massively different depending on the codebase.
Good E2E strongly typed codebase? Can one shot any feature, some small QA, some polishing and it's usually good to ship.
Plain javascript? Object oriented? Injection? Overall magic? Claude can work there but is not a pleasant experience and I wouldn't say it accelerates you that much.
That's essentially the thing we are calling "cognitive debt".
I have a chapter with one small thing to help address that here - https://simonwillison.net/guides/agentic-engineering-pattern... - but it's a much bigger topic and will require extensive exploration by the whole industry to figure out.
Agreed. I've spent the last few years building an EMR at an actual agency and the idea that users know what they want and can articulate it to a degree that won't require ANY technical decisions is pure fantasy in my experience.
For example, "Generate me some repeatable code to ask system X for data about Y, pull out value Z, and submit it to system W."
Doesn't it sound like the "right incantation"? That's the point of LLMs, they can understand (*) intent. You'd get the same result saying "do tdd" or "do the stuff everyone says they do but they don't, with the failing test first, don't remember the name, but you know what I'm saying innit?"
I'm perhaps uncharitable, and this article just happens to take the collateral damage, but I'm starting to see the same corruption that turned "At regular intervals, the team reflects on how to become more effective" into "Mandatory retro exactly once every fortnight, on a board with precisely three columns".
If you accelerate the pace of code creation it inevitably creates bottlenecks elsewhere. Code review is by far the biggest of those right now.
There may be an argument for leaning less on code review. When code is expensive to produce and is likely to stay in production for many years it's obviously important to review it very carefully. If code is cheap and can be inexpensively replaced maybe we can lower our review standards?
But I don't want to lower my standards! I want the code I'm producing with coding agents to be better than the code I would produce without them.
There are some aspects of code review that you cannot skimp on. Things like coding standards may not matter as much, but security review will never be optional.
I've recently been wondering what we can learn from security teams at large companies. Once you have dozens or hundreds of teams shipping features at the same time - teams with varying levels of experience - you can no longer trust those teams not to make mistakes. I expect that the same strategies used by security teams at Facebook/Google-scale organizations could now be relevant to smaller organizations where coding agents are responsible for increasing amounts of code.
Generally though I think this is very much an unsolved problem. I hope to document the effective patterns for this as they emerge.
Weirdly this article doesn't really talk about the main agentic pattern
- Plan (really important to start with a plan before code changes). iteratively build a plan to implement something. You can also have a colelctive review of the plan, make sure its what you want and there is guidance about how it should implement in terms of architecture (should also be pulling on pre existing context about your architecure /ccoding standards), what testing should be built. Make sure the agent reviews the plan, ask the agent to make suggestions and ask questions
- Execute. Make the agent (or multiple agents) execute on the plan
- Test / Fix cycle
- Code Review / Refactor
- Generate Test Guidance for QA
Then your deliverables are Code / Feature context documentation / Test Guidance + evolving your global/project context
I know they won’t stop using AI so giving them a directives file that I’ve tried out might at least increase the quality of the output and lower my reviewing burden.
Open to other ideas too :)
I'm hoping to add more on that topic as I discover other patterns that are useful there.
I’ve always been partial to integration tests too. Hand coding made integration tests feel bad; you’re almost doubling the code output in some cases - especially if you end up needing to mock a bunch of servers. Nowadays that’s cheap, which is super helpful.
That's fine; these boats are coming daily now. I'll catch the next one if I need to.
Always, even before all this madness. It sounds more like a function of these teams CR process rather than agents writing the code. Sometimes super large prs are necessary and I've always requested a 30 minute meeting to discuss.
I don't see this as an issue, just noise. Reduce the PR footprint. If not possible meet with the engineer(s)
Country AND Western!
It sounds like you have a misunderstanding of what LLMs are/can do.
Imagine that you only get one first interaction with a person that you're having try to build something and you're trying to minimize the amount back and forth.
For humans this can be something like an instruction manual. If you've put together more than a few things you quickly realize that instruction manuals vary highly in quality, some will make your life much easier and other will leave you confused.
Lastly, (human) intent is a social construct. The more closely you're aligned with the entity in question the more it's apt to fully comprehend your intent. This is partially the reason why when you throw a project at workers in your office they tend to get it right, and when you throw it towards the overseas team you'll have to check in a lot more to ensure it's not going off the rails.
There's no suggestion that this is The Only Blessed Way.
I.e. have a `planning/designs/unbuilt/...` folder that contains markdown descriptions of features/changes that would have gotten a PR. Now do the review at the design level.
Is the boat:
1) unmissable since the tools get better all the time and are intelligent
or
2) nearly-impossible to board since the tools will replace most of the developers
or
3) a boat of small productivity improvements?
?
I hear what you're saying, but I think it's going to be entertaining watching people go "I guess this is why we paid Bob all of that money all those years".
I don't think we'll get their by scaling current techniques (Dario disagrees, and he's far more qualified albeit biased). I feel that current models are missing critical thinking skills that I feel you need to fully take on this role.
the interesting engineering problem is that the two feedback loops run on different timescales - functional feedback is immediate (did the click work?) but behavioral feedback is lagged and probabilistic (the session might get flagged 10 requests from now based on something that happened 5 requests ago). teaching an agent to reason about that second loop is the unsolved part.
If that's true, then I would think the emphasis in code review should be more on test quality and verifying that the spec is captured accurately, and as you suggest, the actual implementation is less important.
Agree with everything else you said except this. In my opinion, this assumes code becomes more like a consumable as code-production costs reduce. But I don't think that's the case. Incorrect, but not visibly incorrect, code will sit in place for years.
I don't care how cheap it is to replace the incorrect code when it's modifying my bank account or keeping my lights on.
Yea, a big part of my planning has included what verification steps will be necessary along the way or at the end. No plan gets executed without that and I often ask for specific focus on this aspect in plan mode.
The problem is Claude Code has a planning mode baked in, which works really well but is quite custom to how Claude Code likes to do things.
When I describe it as a pattern I want to stretch a little beyond the current default implementation in one of the most popular coding agents.
* no-op tests
* unit tests labeled as integration tests
* skipped tests set to skip because they were failing and the agent didn’t want to fix them
* tests that can never fail
Probably at any given time the tests are 2-4% broken. I’d say about 10% of one-shot tests are bogus if you’re just working w spec + chat and don’t have extra testing harnesses.
They are, and that's deliberate.
Something I'm finding neat about working with coding agents is that most of the techniques that get better results out of agents are techniques that work for larger teams of humans too.
If you've already got great habits around automated testing, documentation, linting, red/green TDD, code review, clean atomic commits etc - you're going to get much better results out of coding agents as well.
My devious plan here is to teach people good software engineering while tricking them into thinking the book is about AI.
It can still cheat, but it's less likely to cheat.
The only problem is... they still take much longer to _run_ than unit tests, and they do tend to be more flaky (although Claude is helpful in fixing flaky tests too). I'm grateful for the extra safety, but it makes deployments that much slower. I've not really found a solution to that part beyond parallelising.
In my experience, this heavily depends on the task, and there's a massive chasm between tasks where it's a good and bad fit. I can definitely imagine people working only on one side of this chasm and being perplexed by the other side.
The claim here is profound: comprehension of the codebase at the function level is no longer necessary
It wasn't profound the first time, and it's even dumber that people keep repeating it - maybe they take all the time they saved not writing, and use it to not read.
A nice way to realize why this AI wave hasn't produced massive economy growth, it is mostly touching parts of economy which are parasitic and can't really create growth.
https://www.slater.dev/2025/09/its-time-to-license-software-...
I don’t think you have to square them because those sentiments are coming from different people. They are also coming from people at different points along the adoption curve. If you are struggling and you see other people struggling at the beginning of the adoption curve it can be quite difficult difficult to understand someone who is further along and does not appear to be struggling.
I think a lot of folks who have struggled with these tools do so because both critics and boosters create unrealistic expectations.
What I recommend is you keep trying. This is a new skill set. It is a different skill set. Which other skills that existed in the past remain necessary is not known.
0: https://www.codewithjason.com/examples-pointless-rspec-tests...
I was expecting tips on code review instead based on your comment and GP.
Many times I've observed that the tests added by the model simply pass as part of the changes, but still pass even when those changes are no longer applied.
> I hear a lot of people saying the same, but similarly a bunch of people I respect saying they barely write code anymore. It feels a little tricky to square these up sometimes.
It squares up just fine.
You ever read a blog post or comment and think "Yeah, this is definitely AI generated"? If you can recognise it, would you accept a blog post, reviewed by you, for your own blog/site?
I won't; I'll think "eww" and rewrite.
The developers with good AI experiences don't get the same "eww" feeling when reading AI-generated code. The developers with poor AI experiences get that "eww" feeling all the time when reviewing AI code and decide not to accept the code.
Well, that's my theory anyway.
I think trying agents to do larger tasks was always very hit or miss, up to about the end of last year.
In the past couple of months I have found them to have gotten a lot better (and I'm not the only one).
My experience with what coding assistants are good for shifted from:
smart autocomplete -> targeted changes/additions -> full engineering
1) Having review loops between agents (spawn separate "reviewer" agents) and clear tests / eval criteria improved results quite a bit for me. 2) Reviewing manually and giving instructions for improvements is necessary to have code I can own
Engineering is the practical application of science and mathematics to solve problems. It sounds like you're maybe describing construction management instead. I'm not denying that there's value here, but what you're espousing seems divorced from reality. Good luck vibecoding a nontrivial actuarial model, then having it to pass the laundry list of reviews and having large firms actually pick it up.
Quite a heavy-lifting word here. You understand why people flagged that post right? It's painfully non-human. I'm all for utilizing LLM, but I highly suggest you read Simon's posts. He's obviously a heavy AI user, but even his blog posts aren't that inorganic and that's why he became the new HN blog babe.
[0]: I personally believe Simon writes with his own voice, but who knows?
I would not equate software engineering to "proper" engineering insofar as being uttered in the same sentence as mechanical, chemical, or electrical engineering.
The cost of code is collapsing because web development is not broadly rigorous, robust software was never a priority, and everyone knows it. The people complaining that AI isn't good enough yet don't grasp that neither are many who are in the profession currently.
Also, the same thing applies to feedback loops with compilers and linters as well: they provide objective feedback that then the AI goes and fixes, verifiably resolving the feedback.
Even with less verifiable things like using specifications, the fact that it relies on less objective grounding metrics doesn't mean there's no change in the model's behavior. I'm sure if you looked at the code that a model produced and the amount of intervention necessary to get there for a model that was asked to produce something without a specification versus with one, you would definitely see an objective difference on average. We're already getting objective studies regarding AGENTS.MD
But it stands to reason that would be a huge shift if a system accessible to non-technical users could mostly handle those edge cases, even when "handle" means failing silently without taking the entire thing down, or simply raising them for human intervention via Slack message or email or a dashboard or something.
And Bob's still going to get paid a lot of money he'll just be doing stuff that's more useful than figuring out how negative numbers should be parsed in the ETL pipeline.
Is your AI not even doing try/catch statements, what century are you in?
That said, we should not underestimate the ability of companies to limp along with something broken and buggy, especially when they're being told there's no budget to fix it. (True even before LLMs.)
If Opus 4.6 had 100M context, 100x higher throughput and latency, and 100x cheaper $/token, we'd be much closer. We'd still need to supervise it, but it could do a whole lot more just by virtue of more I/O.
Of course, whether scaling everything by 100x is possible given current techniques is arguable in itself.
Yea, we'll see. I didn't think they'd come this far, but they have. Though, the cracks I still see seem to be more or less just how LLMs work.
It's really hard to accurately assess this given how much I have at stake.
> and he's far more qualified albeit biased
Yea, I think biased is an understatement. And he's working on a very specific product. How much can any one person really understand the entire industry or the scope of all it's work? He's worked at Google and OpenAi. Not exactly examples of your standard line-of-business software building.
You should be planning out the tests to properly exercise the spec, and ensuring those tests actually do what the spec requires. AI can suggest more tests (but be careful here, too, because a ballooned test suite slows down CICD), but it should never be in charge of them completely.
Problem is, the way I've been trained to test isn't exactly antagonistic. QA does that kind of thing. Programmers writing tests are generally rather doing spot checks that only make sense if the code is generally understood and trustworthy. Code LLMs produce is usually broken in subtle, hard to spot ways.
It does 2 things that are very important, 1: reviewing should not be done last, but during the process and 2: plans should result into verifyable specs, preferably in a natural language so you can avoid locking yourself into specific implementation details (the "how") too early.
Eventually I do think it will be 2.
I think you’ve got to make hay while the sun shines. Nobody knows how this is all going to play out, I just want to make sure I’m at the forefront of it.
Yeah, I'm not sure I agree with what I said there myself!
> Incorrect, but not visibly incorrect, code will sit in place for years.
If you let incorrect code sit in place for years I think that suggests a gap in your wider process somewhere.
I'm still trying to figure out what closing those gaps looks like.
The StrongDM pattern is interesting - having an ongoing swarm of testing agents which hammer away at a staging cluster trying different things and noting stuff that breaks. Effectively an agent-driven QA team.
I'm not going to add that to the guide until I've heard it working for other teams and experienced it myself though!
However a thing I'm finding quite valuable in my own workflows, but haven't seen much discussion of, is spending meaningful time with AI doing meta-planning of that document. For example, I'll spend many sessions partnered with AI just iterating on the draft document, asking it to think through details, play contrarian, surface alternatives, poke holes, identify points of confusion, etc. It's been so helpful for rapidly exploring a design space, and I frequently find it makes suggestions that are genuinely surprising or change my perspective about what we should build.
I feel like I know we're "done" when I thoroughly understand it, a fresh AI instance seems to really understand it (as evaluated by interrogating it), and neither of us can find anything meaningful to improve. At that point we move to implementation, and the actual code writing falls out pretty seamlessly. Plus, there's a high quality requirements document as a long-lived artifact.
Obviously this is a heavyweight process, but is suited for my domain and work.
ETA one additional practice: if the agent gets confused during implementation or otherwise, I find it's almost always due to a latent confusion about the requirements. Ask the agent why it did a thing, figure out how to clarify in the requirements, and try again from the top rather than putting effort into steering the current session.
I do this with code too.
I'm thinking about how to solve the problem and how to express it in the programming language such that it is easy to maintain. Getting someone/something else to do that doesn't help me.
But different strokes for different folks, I suppose.
I think this is the main point where many people’s work differs. Most of my work I know roughly what needs changing and how things are structured but I jump between codebases often enough that I can’t always remember the exact classes/functions where changes are needed. But I can vaguely gesture at those specific changes that need to be made and have the AI find the places that need changing and then I can review the result.
I rarely get the luxury of working in a single codebase for a long enough period of time to get so familiar with it that I can jump to particular functions without much thought. That means AI is usually a better starting point than me fumbling around trying to find what I think exists but I don’t know where it is.
Pretty recently (a couple weeks ago). I give agentic workflows a go every couple of weeks or so.
I should say, I don't find them abysmal, but I tend to work in codebases where I understand them, and the patterns really well. The use cases I've tried so far, do sort of work, just not yet at least, faster than I'm able to actual write the code myself.
it { expect(classroom).to have_many(:students) }
I'm not sure that they're particularly harmful any more though. It used to be that they added extra weight to your test suite, meaning when you make changes you have to update pointless tests.
But if the agent is updating the pointless tests for you I can afford a little bit of unnecessary testing bloat.
And I have the AI deal with "knowing how to do it" as well. Often it's slower to have it do enough research to know how to do it, but my time is more expensive than Claude's time, and so as long as I'm not sitting around waiting it's a net win.
I’m not sure it’s really true in practice yet, but that would certainly be the claim.
I’ve yet to see these things do well on anything but trivial boilerplate.
To answer your question, I’ve tried both Claude code and Antigravity in the last 2 weeks and I’m still finding them struggling. AG with Gemini regularly gets stuck on simple issues and loops until I run out of requests, and Claude still just regularly goes on wild tangents not actually solving the problem.
> smart autocomplete -> targeted changes/additions -> full engineering
Define "full engineering". Because if you say "full engineering" I would expect the agent to get some expected product output details as input and produce all by itself the right implementation for the context (i.e. company) it lives in.
Thats a little harsh. I think most everyone would agree we're in a transformative time for engineering. Sure theres hype, but the adoption in our profession (assuming you're an engineer) isn't waning.
There's no actual way to determine if any words are from a silicon token generator or meat-based generator. It's not AI, it's human! Emdash. You're absolutely right!
system failure.
I think the externalities are being ignored. Having time and money to train engineers is expensive. Having all the data of your users being stolen is a slap in the wrist.
So replacing those bad worekrs with AI is fine. Unless you remove the incentives to be fast instead of good, then yeah AI can be good enough for some cases.
Admittedly, in the absence of halfway competent static type checking, it does seem like a good way to prevent what would be a very bad regression. It doesn’t seem worse than tests which check that a certain property is non-null (when that’s a vital business requirement and you’re using a language without a competent type system).
And the progress is slowing down in such a way, that knowledge learned today will not be outdated anymore?
Should investors be worried, since AGI is not coming anymore?
I'm not sure I agree with this. I don't think there needs to be a whole spec & documentation process before plan mode.
There is alternative thought leadership that the waterfall approach for building out projects is not the right Agentic pattern[1].
Planning itself can be such an intensive process where you're designing and figuring out the specs on the fly in a focused manner for the thing the agent will actually develop next. Not sure how useful it is to go beyond this in terms of specs that live outside of the Agentic loop for what should be developed now and next.
I've evolved my own process, originally from plain Claude Code to Claude Code with heavy spec integrated capabilities. However, that became a burden for me: a lot of contextual drift in those documents and then self managing & orchestrating of Claude Code over those documents. I've since reoriented myself to base Claude Code with a fairly high-level effort specific to ad-hoc planning sessions. Sometimes the plans will revolve around specific GitHub issues or feature requests in the ticketing system, but that's about it.
[1] https://boristane.com/blog/the-software-development-lifecycl...
So, you have a code writing droid that is aligned towards writing good clean code that humans can read. Then you have an implementation droid that goes into actually launching and running the code and is aligned with business needs and expenses. And you have a QA droid that stress tests the code and is aligned with the hacker mindset and is just slightly evil, so to speak.
Each droid is working together to make good code, but also are independent and adversarial in the day to day.
This is exactly the sort of future I'm afraid of. Where the people who are ostensibly hired to know how stuff works, out source that understanding to their LLMs. If you don't know how the system works while building, what are you going to when it breaks? Continue to throw your LLM at it? At what point do you just outsource your entire brain?
The benefit is I can keep some things ticking over while I’m in meetings, to be honest.
Especially good to navigate the code if you're unfamiliar with it (the code). If you have known the code for good, you'll find it's usually faster to debug and code by yourself.
Opus 4.6 with claude code vscode extension
Gemini running an benchmark- everything ran smoothly for an hour. But on verification it had hallucinated the model used for judging, invalidating the whole run.
Another task used Opus and I manually specified the model to use. It still used the wrong model.
This type of hallucination has happened to me at least 4-5 times in the past fortnight using opus 4.6 and gemini-3.1-pro. GLM-5 does not seem to hallucinate so much.
So if you are not actively monitoring your agent and making the corrections, you need something else that is.
Because, after they're done/have finished executing, I guess you still have to "check" their output, integrate their results into the bigger project they're (supposedly) part of etc, and for me the context-switching required to do all that is mentally taxing. But maybe this only happens because my brain is not young enough, that's why I'm asking.
I think it can (and is) shifting very rapidly. Everyone is different, and I’m sure models are better at different types of work (or styles of working), but it doesn’t take much to make it too frustrating to use. Which also means it doesn’t take much to make it super useful.
I.e. the point where the agent writes all the code and you just verify.
We didn't ask if type-based autocomplete was "intelligent" before we started using that.
Treat coding agents as tools and figure out what they can and cannot do and how best to use them.
I think the relative comfort we've enjoyed as software engineers is going to disappear eventually. I just want to be the last to go.
My whole career, I've remained valuable by staying at the forefront of what is possible and connecting that to users' needs. Nothing has changed about my approach from that perspective.
I'm not an investor so I have no idea how they should think.
Because that's what they'll be used for.
Suppose you iterate through many sessions of lightweight planning, implementation, and code review. It _feels_ high velocity, you're cranking through the feature, but you've also invested a lot of your time and energy (planning isn't free, and code review and fit-for-purpose checks, in particular, are expensive). As often happens -- with or without AI -- you get towards the end and realize: there might have been a fundamentally better approach to take.
The tradeoff of that apparent velocity is that _now_ course correction is much more challenging. Those ephemeral plans are now gone. The effort you put into providing context within those plans is gone. You have an locally optimal solution, but you don't have a great way of expressing how to start over from scratch pointed in a slightly different direction.
I think that part can be really valuable, because given a sufficiently specific arrow, the AI can just rip.
Whether it's worth the effort, I suppose, depends on how high-conviction you are on your original chosen approach.
Also instead of just prompting, having it write a quick summary of exactly what it will do where the AI writes a plan including class names branch names file locations specific tests etc. is helpful before I hit go, since the code outline is smaller and quicker to correct.
That takes more wall clock time per agent, but gets better results, so fewer redo steps.
Opus 4.6 has been out for less than a month. If it was a big shift surely we'd see a massive difference over 4.5 which was november. I think this proves the point, you're not seeing seisimic shifts every 3 months and you're not even clear about which model was the fix.
> I think it can (and is) shifting very rapidly.
Shifting, maybe. But shuffling deck chairs every 3 months.
No. The parent comment said I needed a new model, which I've tried. Being told "just try something else aswell" kind of proves the point.
But the most important thing is that they were reverse engineering gcc by using it as an oracle. And it had gcc and thousands of other c compilers in its training set.
So if you are a large corporation looking to copy GPL code so that you can use it without worrying about the license, and the project you want to copy is a text transformer with a rigorously defined set of inputs and outputs, have at it.
Perfect example. You mean the C compiler that literally failed to compile a hello world [0] (which was given in it's readme)?
> What do you consider simple issues?
Hallucinating APIs for well documented libraries/interfaces, ignoring explicit instructions for how to do things, and making very simple logic errors in 30-100 line scripts.
As an example, I asked Claude code to help me with a Roblox game last weekend, and specifically asked it to "create a shop GUI for <X> which scales with the UI, and opens when you press E next to the character". It proceeded to create a GUI with absolute sizings, get stuck on an API hallucination for handling input, and also, when I got it unstuck, it didn't actually work.
[0] https://github.com/anthropics/claudes-c-compiler/issues/1
the actual use cases we see are mostly legitimate automation - QA teams testing geo-specific flows, price monitoring, research pipelines that need to run at scale without getting rate-limited on the first request. the same problem space as curl-impersonate or playwright-extra, just at the session management layer.
could someone use it for spam? technically yes, same as they could with any headless browser setup. but spam operations generally don't need sophisticated fingerprinting - they're volume plays that work fine with basic tools. the people who need real browser isolation are usually the ones doing something that has a legitimate reason to look human.
Hopefully there is some of lint process to catch my human hallucinations and typos.