"In this post, I’ll cover a third, not-so-obvious approach: building ways for the agent to validate more of its own work before a human has to step in. "
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.
This is where most of my productivity gains have come, I have a special harness I move from project to project now that does my testing orchestration, lots of my work day is setting up a prompt or two early and just letting them loop till they return evidence that the feature is working having gone through the big QA loop.
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.
This what hooks[1] are for, except hooks allow specifying criteria in certain conditions (like the agent believing it’s done and ready to hand back to the user) in a manner that the agent won’t just forget about once it’s a few turns deep, and doesn’t require triggering a whole other LLM instance to read some plain text instructions while you hope it interprets them correctly.
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.
Everyone looking into this and other verification should be moving away from long prompts and complex skills, and looking into hooks.
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.
I think pi-subagents (which can form arbitrarily long chains of subagents, with up to 8 in parallel) and Claude Code‘s new workflows feature, are quite convenient abstractions that can be setup quickly.
Isn't this a bit of an incorrect usage of the term "backpressure"?
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.
It is an incorrect use of what was already a flawed metaphor. Pressure is isotropic. Directed pressure makes no sense, like all other fluid analogies in unrelated fields of engineering.
Wait so cross ventilation, where a breeze will flow through a house if windows are open on opposite sides at a much greater rate than if windows are only open on the upwind side… isn’t really a thing?
If the systems invariants are well defined, and a suite of conformance + requirements tests (ensuring invariance is respected) are defined, wouldn't this be a broad - _'base case'_ - approach in general?
A very long post about a simple and very obvious idea with many different implementations.
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", 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.
> all the flows where a model has the initiative are strictly biased towards unwarranted stops
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.
Yeah I don’t really see the backpressure analogy here - it implies that the agent is constantly producing new stuff, which isn’t really possible since the solution is very detailed specs/goals.
That quote shows an utter disregard for basic human decency.
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.
They are probably reacting to the laughable idea that by making PRs 20% better (or whatever), devs will continue to review the code with sufficient rigor to catch even the bugs they're supposedly now preventing. Assuming such rigor was ever present in their work!
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?!
Interesting ideas for generalizing goals to reduce human labor in human <—> agent interactions. That said, maybe it is better to set up customized skills and infrastructure for large projects? At our early stage of trying to capture value of agentic systems, the good ideas in this article might be premature optimization.
I’m willing to be wrong but this industry-wide emphasis on AI creative/coding workflows seems way over-engineered.
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.
LLMs are too flaky for high quality code. On tougher problems it's very common for an LLM to contradict itself and run in circles. It simply doesn't know what the right thing is, but on each turn it is super confident to do the right thing.
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.
For me it's usually that I start with a single agent, but then I won't have anything to do while it is churning and I have other ideas/features that keep building up that I want to do, so I need to scale, and while I'm scaling I need to start to have those workflows, so eventually I end up with many agents, most which are autonomous working on their own worktrees, but I will have a specific agent that I will talk to more iteratively.
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.
Every large project in the coming back to waterfall. While the problems are certainly known and it was ultimately developed as a straw man, everything else ends up working worse. That said, you shouldn't be thinking pure waterfall as it's drawn up as a strawman, but rather a waterfall variation with feedback loops. But in the end, in very, very many cases, you have to know an end date in order to get things done because so many other things depend on you being done at the same time. If something is going to get done sooner you can't use it anyway without all the other pieces.
I suspect that letting agents spin away unattended for long stretches of time will become less and less popular as more and more companies blow their token budgets and start requiring some answers to difficult questions before agreeing to further loose the purse strings.
I agree. I have gotten an incredible amount of work done iterating with 5-30 minute long agent tasks. But it requires I stay engaged, and not go chill on the beach, which I guess is a lot of agentmaxxers’ goal.
interesting idea, unfortunately programming the structure is equivalent (P=NP) to just programming itself. same as TDD.
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.
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.
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.
[1] https://code.claude.com/docs/en/hooks
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.
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.
It comes from previous posts I’ve come across, but I haven’t considered exactly what you mentioned. That’s on me.
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", 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.
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.
https://pura.xyz
https://github.com/puraxyz/puraxyz/blob/main/docs/paper/main...
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?!
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.
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.
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.
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.