The damn thing _talks_. You can just _speak_ to it. You can just ask it to do what you want.
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.
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
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.
> A comprehensive test suite is by far the most effective way to keep those features working.
there is no mention at all about LLMs' tendency to write tautological tests--tests that pass because they are defined to pass. Or, tests that are not at all relevant or useful, and are ultimately noise in the codebase wasting cycles on every CI run. Sometimes to pass the tests the model might even hardcode a value in a unit test itself!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.
- 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.
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.
"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.
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.
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.
For a high level description of what this new way of engineering is about: https://substack.com/@shreddd/p-189554031
The thing I keep wrestling with is where exactly to place those checkpoints. Too frequent and you've just built a slow pair programmer. Too infrequent and you're doing expensive archaeology to figure out where it went sideways. We've landed on "before any irreversible action" as a useful heuristic, but that requires the agent to have some model of what's irreversible, which is its own can of worms.
Has anyone found a principled way to communicate implicit codebase conventions to an agent beyond just dumping a CLAUDE.md or similar file? We've tried encoding constraints as linter rules but that only catches surface stuff, not architectural intent.
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/
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
Running multiple agents concurrently (QA, content, conversions, distribution), we hit this exact wall - agents didn't know what other agents had done, creating duplicate work and missed context.
Solved it with a stupidly simple approach: 1. Single TODO.md with "DO NOW" (unblocked), "BLOCKED", "DONE" sections 2. Named output files per agent type (qa-status.md, scout-finds.md, etc) 3. active-tasks.md for crash recovery - breadcrumbs from interrupted runs 4. Daily memory logs with session IDs for searchability
The key: File-based state is deterministic. After a crash, the next agent reads identical input, same decision rules, same output structure. Zero state collision, zero "what was I thinking?"
Deployment: ~8 agents on cron. They wake, read files, work, write results, die. No persistent terminal. No coordination overhead.
This turned "5 terminal tabs with unmanageable logs" into "grep yesterday's log, see exactly what happened."
Patterns + implementation details: https://osolobo.com/first-ai-agent-guide/
The "give it bash" pattern sounds scary until you realize the alternative is 47 intermediate tool calls that fail silently.
Letting the agent write and run scripts means the agent debugs when something breaks. The feedback loop tightens dramatically.
The trick is sandboxing + cost limits. Not preventing shell access.
Has anyone setup a smooth agent setup for game art assets generation? (AI models already do great for shaders and VFX, but I would really love to automate model + texture + animation pipeline)
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
https://simonwillison.net/guides/agentic-engineering-pattern...
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
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...
- 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”)
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.
Shameless plug: I wrote one. https://marmelab.com/blog/2026/01/21/agent-experience.html
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.
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
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.
I am still not sold on agentic coding. We’ll probably get there within the next couple of years.
Thank you Simon and I'm sure you would quickly fall off from #1 blogger on HN if you did. I insist on this for myself as well.
Somehow we are all getting really good at detecting "written by AI" with primal intuition.
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
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.
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)
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