Backpressure is all you need(lucasfcosta.com) |
Backpressure is all you need(lucasfcosta.com) |
Increased complexity of your systems. Increased pipelines of your system.
You might reduce the likelihood of errors, but at an overproportinal cost of time it takes to complete (which some might argue is irrelevant, but has the cost of human context), and with an way higher time and focus needed for all bugs that the system doesnt work.
You’ll have to fix adapt and maintain all your verification layers, because just because you set them up they are not perfect.
Your testing pipeline becomes incredible slow and you need to maintain it as well.
It’s tremendously weaker than a hands-on approach.
I’ve written this exact same article in January and since then completely switched my position.
Good luck on everyone trying this. You shuffling your own grave and waste time.
this has been an obvious thing to do since at least January (since Geoffrey Huntley published "everything is a ralph loop"), and this is how I've been working: build enough orchestration tooling to be able to automate everything: development container bringup, building it, running the unit tests, doing integration testing, and using the software as eventually an end user. then to iterate set performance goals on an already solid basis so the automated agent ("gym") can go and iterate autonomously, and let you know when it's "done".
I understand this probably does not work if you're on some subscription and not using the API (tokens burn fast), but this has been extremely productive for me.
I've slowly been optimizing for token use through the stack and Claude ends up making very tight for loops for most of the process and keeping token count even lower. It's been nice. A lot of my toil at work is just gone.
"The goal to make longer unattended sessions safe enough to be useful without fully removing the human from the loop. It should also reduce the number of low-quality PRs your teammates have to review for details the agent should have caught itself."
>safe enough to be useful without fully removing the human from the loop
This is the fundamental concept for AI usage, assistance and adoption for every fields not only code generation.
Essentially AI including LLM, ML, DL, is just a tool, like any other automation tools operating based on the principle of expert-in-the-loop as safety and quality gatekeeper, for sensible and responsible decision making [1].
[1] Domain expertise has always been the real moat (brethorsting.com) (519 comments):
with enough scaffolding around self-reflectivity and metrics, it will converge.
Do you have infinite money?
It also presents tradeoffs in compute budget. Cycles spent executing large arrays of tests could mean less tokens spent debugging.
OP quoted the correct definition right at the start:
> In systems engineering, backpressure is the mechanism by which a downstream component signals upstream that it can't accept more work
(the "downstream component" being the human reviewer in this case)
But the measures they propose don't actually do that. They are more like fixed throttle elements which would slow down the rate of submissions of an agent and weed out some low-quality submissions before hitting "downstream".
I'm missing the connection to the actual capacity (or will) that the human developers have to review the submissions.
The three main problems are 1) API usage is deadly expensive 2) Claude is about to make all automation very expensive 3) all the flows where a model has the initiative are strictly biased towards unwarranted stops (checkpointing).
Also, I won't call that "backpressure", there is no producer-consumer disbalance or something similar. From what I can see, the author just proposes a structured feedback loop. That's a discussion about organizational principles for system which consist of multiple unreliable but very complex components and this "backpressure" is just one of the aspects. Personally I find the viable system model framework productive as both a mental model and literal implementation guideline.
Lesser problem is that agent SDKs are bad and building a custom harness is hard.
Ime successful creative execution looks like micro-iterations where each output informs the next creative move.
I can build something incredibly fast from essentially caveman grunt instructions through an LLM harness, iterating as I go.
Optimizing for feeding a huge plan to an agent sounds to me like a net waste of time. And looking over the shoulder of industry peers trying to do this, I don’t see their outputs or throughput some remarkable improvement over what I can produce with minimal fanfare usage.
It absolutely makes sense to have a system in place that allows the code generated by an LLM to be automatically validated but there’s no need to resort to a non-deterministic system for these sort of deterministic pass/fail conditions.
One thing I've been wondering about is how to reliably protect specific portions of the system from unexpected/unnecessary change (for example, a failing test that Claude decides to comment out or rewrite to get it to pass). My only thought for this was to automatically revert test changes during specific portions of the implementation, but that feels overly rigid and potentially prevents things like refactoring code.
- Define the task and the goal, write a short spec document (markdown is fine)
- Point the agent at it in plan mode and have it write the plan to disk with phases. Iterate on its plan if necessary here and now.
- Have each agent tackle a phase and have it update it as a living document (switch models if some phases are more difficult than others)
- Clear and repeat until done
I've never had to overcomplicate this and it's worked both on enterprise-scale projects and personal projects. I am not sure what I'm missing - if anything.
If you put all these checks in your stop hook and your git commit hook, your repo docs can tell your agent that checks will run automatically when it stops work, and it should fix any problems found.
It’s wonderful to reintroduce determinism at the QA end of your process. I find it very calming to know the agent can’t skip or forget to check its work because with hooks the checks are run by the harness.
Oh boy.
It is the responsibility of the person running the coding agent to make sure the resulting PRs are high quality. Putting that on your team mates, or worse, random open source project maintainers on the internet, is the definition of an extractive contribution.
Put another way, who are they supposed to hire to tell these low quality PRs apart from the high quality ones? Who even knows how to do something like that?!
My agent forces this workflow by disabling modifications outside the coding step.
I added looping to this not too long ago. https://github.com/hsaliak/std_slop/blob/main/docs/mail-loop...
This gives me the best of both worlds, hand curated reviews and automation. I often get the best quality if I do both, with an agent doing a pass first.
Next, Vercel, already handle this correctly. It takes special effort to violate "least surprise" here. Cmd-click on a link, should open it in a new tab.
It does appear to be an issue with SimpleAnalytics, now Adobe's,
onclick="saAutomatedLink(this, 'outbound'); return false;"
Good sites, good blogs, *don't override onclick for links.* Or handle it correctly. I'll leave an issue on the github.
Between your footer, and dotfiles repo, OP does seem to appreciate standards & norms, in principle.
The more guardrails you provide the more it cheats.
AI is like a wild animal that needs to do something, and it takes a fair bit of work to corner it. And only when it's cornered and at the point of giving up, can you then offer it a way out.
If you don't do what I said, I can guarantee it's fooling you somehow.
A pre-commit hook has been wonderful. Sure, you can add instructions, but pre-commit hooks are where you want to put the guards.
Called it rik, and it's on GitHub if anyone's interested checking it.
https://github.com/puraxyz/puraxyz/blob/main/docs/paper/main...
as usual, the tool isnt really doing whats listed on its label.
however, people are different so this might improve someones capability to deploy LLMs. might even provide better evidence where actual brain power is needed.
The main kind of pressure I'm feeling is the pressure of the giant AI, GPU & datacenter companies with their insane capital expenditure and circular deals, trying to get enough people to develop an expensive reliance on their service. And the more expensive, the better, so don't just pay for the LLM to code for you, have another LLM interact with the first LLM and pay double, treble, 5x or whatever. Then you can get the most refined slop.
By all means add tons of quality gates to your SDLC pipeline. But thinking about slowness purely for the sake of slowness will not solve your problems.
My gut reaction, as a professional developer, to my (previous) company's AI mandates was an instinctive "wait but..." -- it didn't logic out to me. Now that I have much more AI experience under my belt, I understand the tension, it's a superpower and net-net ok so more features and more "stuff" will be built. But it's a very hard thing to balance. It's always been a bad idea for a company to position themselves as the one with more "stuff" in it.
Fuck, we’re so cooked.
- single-piece flow means not making large batches of things and then sending them all downstream at once, but instead working on one thing at a time so downstream has a chance to reject before too much of the wrong thing is produced.
- autonomation (or jidoka) means giving the machine the ability to detect when something is wrong and not continue at that point.
- poka-yoke is a process that forces results to be conformant by construction.
Any and all of these terms would be better than backpressure in this context.
(This made me realise that lean people have been spending decades dealing with the problems we encounter with the new robots that write code. Half of the lean philosophy is about setting up processes and structures that have positive optionality on people's creativity, without undue requirement on their level of responsibility. That's exactly what we want for robots that write code too. We want to capture the benefits of what they do well, without suffering from their innumerable mistakes. But we can't just chastise them for making mistakes, so we have to think the way lean people do.)
It comes from previous posts I’ve come across, but I haven’t considered exactly what you mentioned. That’s on me.
I think teams need to be able to write nested workflows that transition between code-led and agent-led, with either supporting human-in-the-loop checkpoints.
Been iterating on what this should look like at our startup (https://www.amika.dev/). Model labs are also improving capabilities here, such as Codex's `/goal` and Claude Code's dynamic workflows[1]
The points about API usage cost still stand, but model intelligence is getting cheaper every month! No need to use the frontier model for every part of the work.
/goal is a dynamic workflow itself, from what I know. Dynamic workflows do not hold the initiative (and can't use any libraries or I/O).
Dynamic workflows do not prevent checkpointing.
I don't see the actual point of your startup, it's a cheap idea - such as most LLM startups out there.
I don't see how models are getting cheaper - I clearly see the opposite trend.
Can you elaborate on what you think causes such a bias? My experience is that Qwen3.6, Claude Sonnet 4.6 and Opus 4.6/4.7 will work as far as they can given direction and a way to test their work. My so-far limited experience with Opus 4.8 is that it does stop somewhat earlier for feedback, but in places where I am glad it is checking assumptions or where I agree with it identifying a change in scope (for example, where the following work deserves a separate commit or merge request). I would call those justified stops rather than unwarranted.
You can't express orchestration in terms of "backpressure" only, I think.
Implement-Review-Repeat loop does not involve backpressure in the strict meaning of the term.
Maybe I've chosen hardmode to learn C with LLM assistance, plus my pet project turned out to be a bit less trivial then anticipated. But I know that I have to think three times about my choices how to deal with C problems and seeing how a LLM struggles to give reasonable answers is a a huge red flag and forces me to think about it a fourth time.
Doing all this with a fast autonomous workflow with just little user guidance is asking for trouble.
I suspect that the “right” way to use LMs in coding, including accounting for focus, control, and costs is not a settled debate. We probably haven’t even seen the best ideas yet. But I’m really dislike the maximalist approach.
May you speak a little more about how you're approaching this? I was thinking of doing similar
So e.g. I may have 1 agent that I ask and iterate on with directly, and 9 agents that work separately on their own.
I will utilize this 1 agent on features I care most about and want to guide and iterate on in as much detail as possible.
it works.
It's here to stay, and IMO once VLA-driven robots enter the real world there will be enough money to pay for the datacenters. This coding stuff is great but there are only so many engineers to sell to.
Satya Nadella once said (more or less) "if AI is so good, why doesn't it show up in GDP?"
That's gonna be the step where it shows up in the GDP. Being able to train a machine to solve any problem that can be phrased in tokens (i.e.: most of them) is going to remake society.
The progress is bottlenecked by how things are done with C. Most features went through an considerable LLM research phase to find the sweet spot how I want to solve it. In the feature phase I've coded more and more on my own and used the LLM as a reviewer on the critical parts.
I know a few things about C, but my knowledge is very passive, but there is already a path carved in my brain from influential C programmers you find on yt and blogs. That being said, LLMs will actively gatekeep and bullshit you about C and system architecture. You need to drill down hard, from different angles i.e. what is efficient and what is pragmatic or you question about how hardware actually works. They still may have blind spots and casually keeping things from you. Naively asking for code snippets will be tutorial style and often backfires, not unseen that they rant about their own code in another session. So you need to maintain a healthy balance of knowledge from trustworthy humans and the quickly available, but potentially inaccurate/incomplete LLM knowledge. Be critical, you are the captain.
LLMs are not bad at reviewing C code. They catch a lot of noob mistakes and are more helpful then compilers. But again they tend to produce tutorial solutions. They litter everything with malloc, which is something I am trying to avoid. I've not touched arenas yet, I want to fall flat on raw C to value them. They can teach you arenas, but I suggest to cross check with the real world, especially because you want ergonomics and correctness on that behalf.
Most C adjacent topics are easy to pick up on the go. Build systems, debugging, API vs ABI, static vs dynamic linking, macros, etc. A seasoned programmer can get a quick overview and ad hoc help, for things that are often buried in mediocre docs.
Also if you happen to come from more higher level languages or dynamic languages. Expect C to require easily 5x more code for literally anything.
If you prefer to learn the language more structured, work your way towards learning how to create data structures and algorithms in C. Make a clear distinction between static and dynamic allocation and learn along the road how C/hardware wants you to deal with memory.
I post this angry comment because LLMs are colonizing the language we use for creating an earnest and genuine tone in online discussion and I sometimes wonder if the suspicion surrounding LLM-ish language is worse for the health of our online spaces than the LLM slop itself. Thinking about it, I don't think it is; and it would be impossible to measure anyway.
On checkpointing: I explained myself poorly. You're right that using higher level workflows doesn't turn off checkpointing. One can simply make harnesses non-interactive, but that can make models lose coherence over long tasks (because they can't ask for feedback). A higher level coordinator (/goal, CC dynamic workflows) is designed to provide this feedback without human intervention.
On price: older models keep getting cheaper, and most tasks don't need frontier capability. (I'm ignoring the part about subscription subsidies right now, and just talking about API price for tokens)
On my startup Amika: we run programmable cloud computers for agents, plus the workflow systems to guide them. We let people run any agent (Codex, Claude, etc.), prompt it from anywhere (Slack, web, CLI + SSH, API). It's like devboxes for humans + agents, with guardrails[1] to deterministically ensure things about the changes coding agents make (ie don't let agent modify module boundaries, require every DB query carry a multi-tenant org ID filter).
Maybe our website is bad at explaining it, in which case I appreciate any feedback!
And
> he's proposing a method for how to do so using agents
Are not in agreement. The claim being made is that you shouldn't be sending PRs you haven't personally vetted to be high quality. Definitionally a bot cannot be used to personally vet something.
I am not asking about time or completeness. I am asking if this person is spending 1 dollar to make more than a dollar, or if they are spending 1 dollar to make less than a dollar.
Any other criteria is not necessary to consider, if the activity is not profitable.
you can't just handwave this all away with "vibe baby, vibe" and then high-fiving each other that oh you're so clever because you manually write code/think the code is too high.
Another thing I have in the general sdlc process is having it add enough logging to verify features are turned on, configured as we expected, and that becomes enough feedback for most of my features.
I've been mostly focusing on being able to replicate this across stacks greater than 3 projects so far (with the eventual goal of having an agent be able to orchestrate our complete infra stack, and this being a large component of a DR plan to rebuild).
None of this is really new for us, I'm just the most knowledgeable in my group in how the different products across teams glue together so I've been creating these rube goldbergs as a prototype, and then having it iterate on codifying the parts that don't need a constant LLM. We were blessed to have an engineer a decade ago build out tooling for local container automation that matches 95% of the deployed infra stack. That last 5% sucks when you fall into it, but that's always been a truth. I've added and expanded the tool over the years with making it act more like the deployed environment networking wise, but a lot of things don't end up working well in docker containers on M series macs when most of our complicated virtualization in our private cloud can't run on them yet...
to give you an example just recently I've coded a feature that for our shuffle operation can report which channel did the bytes flow through (as the PR giving us the plumbing underneath has landed upstream recently). what this basically means is that you run the shuffle, you know you've shuffled X bytes (because you have stats on both ends) and then you need to attribute them to different layers. on the first iteration, the count was off. the agent went, debugged, fixed, iterated, and then it was 1.5% off. again, it went, iterated, ... and now we're fine.
part of the task description was that the breakdown must match the known amount of bytes we're shuffling, so the agent took this upon as a self-verification point. so besides running our normal, boring unit tests, integration tests and end-to-end verification harnesses (which it not only has programmatic/cli/API access, but are documented in .md files for projects), it could use this criteria on top to verify.
looking at /usage, my API duration was 2h 43m, and on top of that:
claude-haiku-4-5: 2.7k input, 115.3k output, 16.3m cache read, 867.9k cache write ($3.30)
claude-opus-4-8: 46.9k input, 555.0k output, 166.6m cache read, 2.9m cache write ($115.77)One thing however is a test suite is not usually exhaustive in the sense that any code that passes the tests is valid. Usually tests are more complimentary in nature. Therefore you could still possibly get code degradation, potentially.
Not in the world of AI - if your tests don't catch any known issues, the problem is the tests aren't comprehensive enough. There's no excuse at this point not to have an incredibly comprehensive test suite, to go with your other agent feedback loop constraints
As a software developer, you must never subject your team mates to a PR that you yourself believe to be low quality. The point of code review by others is to catch things that you missed.
There are multiple lines of defense for quality. Yes, automation can and should be one of them, but your own self-review always has to come before review by your team mates.
However, from an operations perspective, a dev is a piece of the qa pipeline with a nonzero error rate, and an optimal throughput rate, above which that error rate rises dramatically.
As a dev, you'll never merge a bad PR; in ops, we want to help you with that goal, and also have plans for what happens when it fails.
Maybe I misunderstand but this seems like a fairly low bar in the test suite only covers existing bugs.
I'd argue that if you aren't going to look at the code you actually need a fully comprehensive test suite - in the sense that if the tests pass, the code is correct and you don't have to look at it at all. The problem is, that isn't very quick to create it seems. Of course, if there is a way to do it quickly in a way that is reproducible by others I'd love to hear about it.
So if you / team are going to implement a new feature, what does that look like? Do you write Gherkin or similar, unit tests or both? Can you provide an example of what that might look like? How much of this has changed for you since the pre-AI days?
You can find some open source examples on github, either directly https://github.com/pgdogdev/pgdog/commits/main/?author=jagge... or through my profile - that repo has a pure-sql integration suite I wrote essentially entirely with AI: https://github.com/pgdogdev/pgdog/tree/main/integration/sql
There's also older work on github you can see over the years, a mishmash and grab bag, I would prefer if more of my work were open source but somehow most employers still default to closed source
Edit: While I'm thinking about it, the other thing you can do with AI is demand that it TDD things - I'm more of a "test all the fucking time" adherent, I don't care whether the tests are written first, but AI is perfectly happy to skate by making a tautological test unless you make it write the test first, ensure it fails correctly, make your change, and don't let it modify the test.