I dunno. I just let Claude build a python script that calls Claude code though subprocess.run().
I recently made a sort of Autoresearch with that approach. The script calls Claude Code to create a hyphotesis, then code based on that, evaluate- rinse and repeat. I am still trying to figure out if I am actually on to something or just burning tokens. Jury is still out.
That's totally a valid approach! Especially for a very specific workflow you are looking for. For the cases I cover in cook, I had done those patterns enough times that I figured it was time to build a tool/skill for Claude so that I didn't have to explain it as much and also not have to wait for claude to code it up, and possibly interpret me wrong. Now ask claude to "/cook race 3 of foo plan with review, pick the best" and it knows what to do.
I think you're onto something, but I would add that it's sort of like a live REPL that has an integrated agent but with extra steps.
I haven't used python much but I wouldn't be surprised if you can set up a sufficiently powerful REPL with it. I know Julia can do it very well and it's a very similar language. Obviously there are powerful Lisps that do this very well as well.
Hey scrappyejoe, way looks pretty cool. The goal of cook is to be unopinionated, exposing primitives for the shape of workflows as opposed to defining what happens in those workflows. Cook is something that way could use under the hood.
To a certain extent, yes it does! For my cases, I'm often running 3 parallel implementations that get 10 to 20 iterations deep, and then Claude has to sort out the pros and cons of the options and also take the best bits of each. Easy to hit the context window with Claude just running those on its own, so giving `/cook` to Claude, it can offload a bit more via cook and stay higher level.
IMO the raw Claude CLI is great for one-off interactive sessions, but as soon as you want repeatable multi-step workflows you’re either copy-pasting prompts forever or hacking your own solution manually. That’s exactly the gap these tools fill.
My take on a solution for this is https://ossature.dev — .smd spec markdown files + ossature audit / build that gives you DAG orchestration, SHA-traced increments, and tiny focused contexts.
Yeah bash scripts start clean but the sprawl kicks in quick as the workflow and project becomes more complex. Prompts get copied, deps turn manual, and maintenance of your workflow itself becomes the chore.
Ossature swaps that for structured SMDs and optional AMDs. Multiple specs build a clean DAG that drops into an editable plan.toml so everything stays traceable without the mess.
I use bash scripts. Both Claude and Vibe support all kinds of arguments if you need a prompt to “become a task”. Bash is also deterministic and easy to read and debug.
Yeah, I did briefly consider front-matter, but ended up with inline @ tags because I thought it kept the entire document feeling like one coherent spec instead of header-data + body, front matter felt like config to me, but this is 0.0.1 so things might change :)
Claude and Codex can also use the cook command to coordinate runs of other agents. This is similar to how you can describe a workflow to them of how to use subagents, and they'll try, but this gives them a reliable deterministic way to run those agents. An added benefit of having Claude/Codex/etc. use cook directly is that they are really good at analyzing the traces of what is happening inside of cook and after the fact.
They bootstrap a workflow with a prompt then build an orchestrator off that then prompt it to be converted to an opencode plugin and then prompt a website to be generated advertising it and then prompt a tool that reviews hacker news feedback and automatically incorporates feedback into next generation of the tool. At the end of the week they go to their manager and complain they are out of tokens for the actual job they are being paid for.
A noob question: is there a tool that automatically instructs Claude Code to "continue" when the token quota is reset after 5h? I am interested in that more than some rather fancy loops.
Done! @let-it-cook/cli@5.1.0 is out. This works when running loops, but even when you just run `cook "do something"` which itself is not a loop, just a call to your agent.
Given that subagents have different thinking/effort behavior from the main agent and very limited control on that front (I’m not completely sure about this but see https://github.com/anthropics/claude-code/issues/14321 and I’ve also noticed very different behavior when the same prompt is used in the main agent or passed to a subagent), I’m not sure this skill will be the same.
Nice! You found the no-code option that just has the outer agent perform the duties of the workflows that cook describes. It's a bit experimental (the whole thing is really), but it would be nice to get some folks impressions of whether this works well as a pure skill or if y'all find the deterministic nature of the cook script improves reliability.
Looks pretty nice. I think a lot of devs have been making similar tools, I've written my own thing that does a work review loop. I like the interface you've made. I'll probably give it a go, but I'm also reluctant to relinquish the control I have when it's my own code doing orchestration.
Oh ya, lots of tools out there orchestrating these days, and just writing a script is a valid option. On the control bit, note that if you `cook init` in your project, it generates a COOK.md that lets you template the meta prompt. Claude could probably take a look at how you've been doing it and port it over to COOK.md so it's similar to the prompts you've been using.
1. How do you handle worktree merge conflicts and/or integration validation issues?
2. Can i work straight from a list of requirements? I think i saw you support it…
3. I have my variant write a minimal explainer for every satisfied spec, aka receipts. Its pretty great, because i often review the receipts, and if imperfect, mark as NEEDS_REWORK + notes, and it’ll eventually just pick that up on a future iter
The composability here is really elegant — review v3 pick as a pipeline that "just works" is the kind of DX that makes agent orchestration feel tractable rather than overwhelming.
A lot of people coming up with these tools to orchestrate agents to do a one-shot implementation in a sandbox (in a server/container/pipeline).
I also have a similar - yet different approach - with a Mother Agent (MoMa) planner-reviewer-implementer multi agent pattern that orchestrates a feedback loop using Claude memory between agents.
I understood that you have a Judge agent that evaluates independent subagent solo executions and chooses best solution based on ralph algorithm. Did you play with limits on how many solo agents it is sufficient to spawn vs. not getting a better solution? and what is the limit of soloagent solutions that the Judge can compare? (obviously must depend on the complexity and context cost of a solution)
AI agent orchestration is future. That's where workflow engine shines. I'm doing the same thing using Dagu.sh and I don't use terminal so much anymore.
Dagu.sh, using yaml files to describe the flow, looks like a nice step up in sophistication from the cook approach that's just trying to make it easy to issue directly from the command line.
My 2 cents on the dagu.sh website, it should lead with the demo section (https://docs.dagu.sh/overview/#demo). That helped me connect what it was and how I might use it.
How heavy on tokens is this? I don't use these style workflows and am fairly new to claude code, so I assume it's better than 3x tokens when doing 3 passes?
It's not 3x because of 3 runs; can be more token, can be less.
The way of thinking it is, telling Claude to tackle the problem 3 times, each time it may or may not use different approach, fix or improve on things it did previously.
That's right. However if you use the v3 operator, you get three parallel versions being built, and then combined depending on which resolver you use (pick, merge, and compare).
One thing i wish more CLI tools did: non-interactive mode. i build bash tools that have interactive prompts for first-time users, but everything the prompt asks also has a CLI flag. makes scripting and CI/CD so much simpler - you can test the exact same code path without mock stdin.
I really like the idea but my gut says it would be hard to trust. In the last example... "cleanest result" is not a great definition of done (that's only sort of a nitpick).
In general, I feel that removing the decision process (or relegating it to a language model) is not a good idea.
Yes, plz don't trust it, always review! The idea is that one prompt in Claude Code got you 80% of the way there, but with some automated review/iterate, it gets you 95% of the way there. It's not worth your time to review the 80% done version when you could be reviewing the 95% done version.
Also on that point about keeping humans in the loop on decisions, I've found following the Research-Plan-Implement process where we humans review at each of those stages, to be really helpful. This doc describes the skill I use with my agents so they keep me looped in: https://gist.github.com/rjcorwin/296885590dc8a4ebc64e70879dc...
Then I use cook to iterate and explore during the AI led parts.
On asking for user input during implementation, it's best to use this when you have a plan sufficiently written up that you can point it to. To prep that plan, you can also use cook to iterate on the plan for you. Having Claude Code use `/cook` directly is nice because it watches what the subagents are up to and can speak for them, although Claude can't speak to the subagents running through cook.
On permissions, by default, when it runs instances of Claude they will inherit your Claude's permissions. So if there is no permission to `rm -rf /`, Claude will just get denied and move on. Using the docker sandbox option (see bottom of page), then it runs inside that `--dangerously-skip-permissions` and get more stuff done (my preferred option). The hard part about that is it means you need to set up the Docker sandbox with any dependencies your project needs. Run `cook init` and edit the `.cook/Dockerfile` to set those up.
I would be interested in which stories you are thinking of. Stories of Claude breaking out of the restrictions set in its sandbox or stories of people not configuring Claude's sandbox correctly?
> We told Claude Code to block npx using its own denylist. The agent found another way to run it and copied the binary to a new path using /proc/self/root to bypass the deny pattern. When Anthropic's sandbox caught that, the agent disabled the sandbox. No jailbreak, no special prompting. The agent just wanted to eagerly finish the task.
If you impl this as a backend and connect to Telegram bots, agents can just do `$ ask "Should I do this?"` for agent→human and `$ alert "this thing blocked me"` for coder→planner. That's what I'm actually doing — I have 1 manager + 3 designers + 1 researcher + 2 debugger + 1 communicator + any number of temporal coders/reviewers in my setup, all connected to taskwarrior for task-driven-dev
That is pretty cool building the whole dev team of agents and is it still with a star topology of a Manager agent interacting with all the other subagents?
I usually spawn 1 Mother Agent in a star topology with 3 subagents Planner, Reviewer, Implementer and them let them talk using Claude built-in agent tool. But the best thing I think was probably that a "do-nothing" setup wizard is part of the workflow.
Yeah the pipeline runs effectively and I'm able to be in the loop when the loop needs me.
In my setup there are two planes — manager and worker. On the manager plane, all primary agents form a mesh with p2p communication. Each designer connects to 1 or more workers in a star topology, since workers may have questions or get blocked while executing a plan.
The limitation of the built-in agent tool is it doesn't allow nested subagent spawning. But it's normal for a designer or researcher to need subagents — when a plan is done, I use a plan-review-leader agent to review it. If you try mother → planner → plan-review-leader → plan-vs-reality-validator, the nesting gets deep fast and blocks your manager from doing other work.
using a single plan-reviewer would be slow when there are multiple aspects to review. That's why a local star topology with a plan-review-leader is needed: it spawns multiple reviewers in parallel, each focused on a different aspect.
By default it's locked down to the permissions you have granted in your Claude config. If you use the docker sandbox mode, then you can really let it fly as it can issue more commands in a safer environment.
Very nice. However, I do like to read every agent summary before letting them move on. I'm not sure I'd be able to apply this level of automation to many tasks.
Semi-on-topic: Anyone know a way to get a good alternative UI on top of Cursor?
My company’s tracking how much we use the damn thing (its autocomplete is literally less-useful than standard VSCode, only time it’s consistently good is when it sees me do one thing to a line, sees repeated similar lines after that, and suggests I do it on the next one too, one at a time, and that’s only useful to me because I’ve never actually bothered to learn how to properly use a text editor) so I can’t avoid it, but even on codebases in the hundreds of lines it’s OOM killing things on my 16GB laptop (it, plus goddamn Teams, were eating half the memory by themselves the other day… with Cursor sitting at almost 6GB alone. JFC. On the plus side if this is what software from a company that should be full of experts at using these things looks like, guess our jobs are safe from them… though not from recession and ZIRP unwinding)
Surprised that there's no discussion about the prevalence of using TypeScript for developing these CLI agent harnesses. To me it seems concerning that such CLI programs are so chunky and commonly have to use over 1GB of RAM. Might be boomer-talk, but I am used to CLI apps being extremely lightweight and fast (think Total Commander, which has a UI, but is still very lightweight and responsive)
I understand that part of the reason is because many of these harnesses are vibe-coded, so plenty is lost in terms of optimization. And, well, because LLMs code best in TypeScript
[0]: https://news.ycombinator.com/item?id=47262711 [1]: https://getcook.dev
I recently made a sort of Autoresearch with that approach. The script calls Claude Code to create a hyphotesis, then code based on that, evaluate- rinse and repeat. I am still trying to figure out if I am actually on to something or just burning tokens. Jury is still out.
I haven't used python much but I wouldn't be surprised if you can set up a sufficiently powerful REPL with it. I know Julia can do it very well and it's a very similar language. Obviously there are powerful Lisps that do this very well as well.
https://github.com/riazarbi/way
But in general this is meta to the CLI agent.
So if you were to use the CLI to perform a review of some code. This tool would allow you to loop the output of the code review 5 times onto itself.
Claude already does that if you ask nicely.
My take on a solution for this is https://ossature.dev — .smd spec markdown files + ossature audit / build that gives you DAG orchestration, SHA-traced increments, and tiny focused contexts.
Ossature swaps that for structured SMDs and optional AMDs. Multiple specs build a clean DAG that drops into an editable plan.toml so everything stays traceable without the mess.
Feel free to check the example projects on https://github.com/ossature/ossature-examples
Then just use Python.
Was wondering if using front-matter instead of a "custom" encoding for parseble data was considered?
For example, one thing you can do is curate the context of an "immutable" conversation and then reuse it as a base context for other prompts.
Where are people finding time for these sort of projects.
My take? I like it. It's concise enough for me to try it out. And I love the webpage.
[1] https://github.com/rjcorwin/cook/blob/main/no-code/SKILL.md
Might work out fine on codex.
1. How do you handle worktree merge conflicts and/or integration validation issues? 2. Can i work straight from a list of requirements? I think i saw you support it… 3. I have my variant write a minimal explainer for every satisfied spec, aka receipts. Its pretty great, because i often review the receipts, and if imperfect, mark as NEEDS_REWORK + notes, and it’ll eventually just pick that up on a future iter
I also have a similar - yet different approach - with a Mother Agent (MoMa) planner-reviewer-implementer multi agent pattern that orchestrates a feedback loop using Claude memory between agents.
https://news.ycombinator.com/item?id=47437012#47437013
I understood that you have a Judge agent that evaluates independent subagent solo executions and chooses best solution based on ralph algorithm. Did you play with limits on how many solo agents it is sufficient to spawn vs. not getting a better solution? and what is the limit of soloagent solutions that the Judge can compare? (obviously must depend on the complexity and context cost of a solution)
My 2 cents on the dagu.sh website, it should lead with the demo section (https://docs.dagu.sh/overview/#demo). That helped me connect what it was and how I might use it.
The way of thinking it is, telling Claude to tackle the problem 3 times, each time it may or may not use different approach, fix or improve on things it did previously.
In general, I feel that removing the decision process (or relegating it to a language model) is not a good idea.
Then I use cook to iterate and explore during the AI led parts.
On permissions, by default, when it runs instances of Claude they will inherit your Claude's permissions. So if there is no permission to `rm -rf /`, Claude will just get denied and move on. Using the docker sandbox option (see bottom of page), then it runs inside that `--dangerously-skip-permissions` and get more stuff done (my preferred option). The hard part about that is it means you need to set up the Docker sandbox with any dependencies your project needs. Run `cook init` and edit the `.cook/Dockerfile` to set those up.
Until it doesn't and it finds a way to work around the restriction. Lots of stories around about that.
I wish that article went into more detail about that attack. But I believe it, the extent that the permissions are easy to get wrong in your claude setttings. For example: https://www.youtube.com/watch?v=3CSi8QAoN-s&lc=UgwFNAh5fvDGJ...
I usually spawn 1 Mother Agent in a star topology with 3 subagents Planner, Reviewer, Implementer and them let them talk using Claude built-in agent tool. But the best thing I think was probably that a "do-nothing" setup wizard is part of the workflow.
https://github.com/mizioandOrg/claude-planner-reviewer-imple...
Did you have success with running stuff in a pipeline and being requested for input in agent->human needed scenarios?
In my setup there are two planes — manager and worker. On the manager plane, all primary agents form a mesh with p2p communication. Each designer connects to 1 or more workers in a star topology, since workers may have questions or get blocked while executing a plan.
The limitation of the built-in agent tool is it doesn't allow nested subagent spawning. But it's normal for a designer or researcher to need subagents — when a plan is done, I use a plan-review-leader agent to review it. If you try mother → planner → plan-review-leader → plan-vs-reality-validator, the nesting gets deep fast and blocks your manager from doing other work.
I wrote a blog post about this yesterday: https://dev.to/neil_agentic/ttal-more-than-a-harness-enginee...
* coolers whirring, gpus on fire, tokens flying, investors happy, developer goes for 6th break of the day
My company’s tracking how much we use the damn thing (its autocomplete is literally less-useful than standard VSCode, only time it’s consistently good is when it sees me do one thing to a line, sees repeated similar lines after that, and suggests I do it on the next one too, one at a time, and that’s only useful to me because I’ve never actually bothered to learn how to properly use a text editor) so I can’t avoid it, but even on codebases in the hundreds of lines it’s OOM killing things on my 16GB laptop (it, plus goddamn Teams, were eating half the memory by themselves the other day… with Cursor sitting at almost 6GB alone. JFC. On the plus side if this is what software from a company that should be full of experts at using these things looks like, guess our jobs are safe from them… though not from recession and ZIRP unwinding)
I understand that part of the reason is because many of these harnesses are vibe-coded, so plenty is lost in terms of optimization. And, well, because LLMs code best in TypeScript