Has the author (not OP) written anything on this topic themselves? This is a blunt comment, because I am fed up with being asked to read LLM content that the prompter thinks is novel and worthwhile because they don't know better.
I can forgive (even root for) someone who puts in the effort themselves to understand a problem and write about it, even if they fall short or miss. They have skin in the game. I have little patience for someone who doesn't understand the disproportionate burden generated content places on the READER.
I can certainly tell they've put the model through the ringer to be terse and use simple language, etc. But I am struggling to separate the human ideas from the vibed ones, and the tone of the whole thing is the usual LLM elevator pitch with "hushed reverence" * "movie trailer cadence".
But "spawn/fork" is just a different way of labeling the fairly-well-understood tactic (I won't call it a strategy) of just how much context to provide sub-agents. Claude Code already "spawns" everytime it does an explore. It can do this concurrently, too.
Beyond that, they seem to express wonder at how well models can use tools:
> In the example above, the agent chose spawn for the independent research tasks and fork for the analysis that needs everything. It made this choice on its own — the model understands the distinction intuitively.
Emphasis mine. They (or the model whose output they blindly published) are anthropomorphizing software that is already designed to work this way. They gave it "fork" and "spawn" tools. Are they claiming they didn't describe exactly how they were supposed to be used in the tool spec?
The criticism about the labeling is valid and I oversold. For clarity, this is what the agent sees:
`spawn`: "Create a spawned child node under your node."
`fork`: "Create a forked child node (inherits parent context) under your
node."
The novelty is less about the distinction between the two, it's the tree generation. I would have served you better, if I just left out the parts that aren't critical to the novelty. Thank you for taking the time to comment.
In all honesty, "would have written it myself but I was too eager to get it out the door" doesn't really make sense to me. You're acknowledging that you took a shortcut to get it out the door (blog post as tech debt is a new one!) - does that mean you'd like to write something up yourself eventually?
I hope so, and would like to read it. In particular, since this is presented as research, I'd be very interested to read about your experimental observations that show the risks/costs/edge cases/limits of this pattern. As it stands, there's no indication of self-critique or really any process of systematic evaluation of the solution.
I enjoyed this! Thank you for taking the time to write it. I like it because I identify with your experiences way more than I do with the standard AI braggy boilerplate. Cheers.
And I deserve and accept the gentle snark at the beginning. I will be sure to let you know when I post something, I'll take your notes :)
on Wednesday, I had a dream about agents. Thursday evening, I talked to Claude about trees. That same night, I pushed out the post. There wasn't much rigor involved but yes I will explore more and report back to you!
You should also try to make context query the first class primitive.
Context query parameter can be natural language instruction how to compact current context passed to subagent.
When invoking you can use values like "empty" (nothing, start fresh), "summary" (summarizes), "relevant information from web designer PoV" (specific one, extract what's relevant), "bullet points about X" etc.
This way LLM can decide what's relevant, express it tersly and compaction itself will not clutter current context – it'll be handled by compaction subagent in isolation and discarded on completion.
What makes it first class is the fact that it has to be built in tool that has access to context (client itself), ie. it can't be implemented by isolated MCP because you want to avoid rendering context as input parameter during tool call, you just want short query.
depends_on is also based on context query but in this case it's a map where keys are subagent conversation ids that are blockers to perform this handed over task and value is context query what to extract to inject.
The actual post and this comment shows how early we are when simple and obvious ideas look novel when first conceptualizing them. Nothing against these ideas though, they are indeed good.
Thank you for the suggestion, I will explore this in the next iteration. I'm learning how to translate how humans do context management into how agents should do them
This kind of research is underrated. I have a strong feeling that these kinds of harness improvements will lead to solving whole classes of problems reliably, and matter just as much as model training.
Imposing a strict, discrete topology—like a tree or a DAG—is the only viable way to build reliable systems on top of LLMs.
If you leave agent interaction unconstrained, the probabilistic variance compounds into chaos. By encapsulating non-deterministic nodes within a rigidly defined graph structure, you regain control over the state machine. Coordination requires deterministic boundaries.
Honest question, have you actually run a DAG-based agent system past like 8 nodes? In my experience the topology isn't what gives you reliability, it's the error handling at each node. You can have the most beautiful DAG in the world and it still falls apart when node 3 returns something subtly wrong and nodes 4-8 all proceed as if everything's fine. The structure gives you the illusion of control.
> This made sense when agents were unreliable. You’d never let GPT-3 decide how to decompose a project. But current models are good at planning. They break problems into subproblems naturally. They understand dependencies. They know when a task is too big for one pass.
> So why are we still hardcoding the decomposition?
Sure, decomposition is already in the pre-training corpus. and then we can do some "instruction-tuning" on top. This is fine for the last mile, but that's it. I would consider this unaddressed and after with the root comment.
Not exactly a surprise Claude did this out of the box with minimal prompting considering they’ve presumably been RLing the hell out of it for agent teams: https://code.claude.com/docs/en/agent-teams
Every time i see some new orchestrator framework worth more than a few hundred loc i cringe so hard. Reddit is flooded with them on the daily and HN has them on the front page occasionally.
My current setup is this;
- `tmux-bash` / `tmux-coding-agent`
- `tmux-send` / `tmux-capture`
- `semaphore_wait`
The other tools all create lockfiles and semaphore_wait is a small inotify wrapper.
They're all you need for 3 levels of orchestration. My recent discovery was that its best to have 1 dedicated supervisor that just semaphore_wait's on the 'main' agent spawning subagents. Basically a smart Ralph-wiggum.
The tmux + lockfile approach is underrated. We went through a whole phase of building proper orchestration infra and ended up ripping most of it out. The overhead of coordinating agents through a framework is often worse than just letting them talk through the filesystem. The dirty secret of multi-agent systems is that the coordination layer is usually where the bugs live, not the agents themselves.
Historically, Claude code used sequential planning with linear dependencies using tools like TodoWrite, TodoRead. There are open source MCP equivalents of TodoWrite.
I’ve found both the open source TodoWrite and building your own TodoWrite with a backing store surprisingly effective for Planning and avoiding developer defined roles and developer defined plans/workflows that the author calls in the blog for AI-SRE usecases. It also stops the agent from looping indefinitely.
Cord is a clever model and protocol for tree-like dependencies using the Spawn and Fork model for clean context and prior context respectively.
I have yet to read this article (in full), but I love trees! As an amateur AST transformation nerd. Kinda related but I’ve been trying to figure out how to generalize the lessons learned from this experiment in autogenerating massive bilingual dictionary and phrasebook datasets: https://youtu.be/nofJLw51xSk
Into a general purpose markup language + runtime for multi step LLM invocations. Although efforts so far have gotten nowhere. I have some notes on my GitHub profile readme if anyone curious: https://github.com/colbyn
(I really dislike the ‘agentic’ term since in my mind it’s just compilers and a runtime all the way down.)
But that’s more serial procedural work, what I want is full blown recursion, in some generalized way (and without liquid templating hacks that I keep restoring to), deeply needed nested LLM invocations akin to how my dataset generation pipeline works.
PS
Also I really dislike prompt text in source code. I prefer to factor in out into standalone prompt files. Using the XML format in my case.
Claude basically does this now (including deciding when to use subagents, tools, and agent teams). I built a similar thing a month ago and saw the writing on the wall.
This is the interesting constraint though. Have you found going deeper than 2 levels actually helps? Every time I've let agents spawn recursively the quality at depth 3+ drops off a cliff. Not because the model is dumber, the accumulated context drift means the deep child is solving a slightly different problem than what the root intended. Curious if your tree structure has a natural depth limit in practice or if it scales.
Are you sure? In Opencode they can, but it's hard to track them then (say, if you want to steer them) -- you have to click through. It would be nice to have a dynamic execution graph alongside the text.
Yeah exactly. I noticed Claude start doing exactly this a month ago too. It recursively breaks problems down while allowing you to either change direction at each level or keep going. This is where claude jumped up to be legitimately better at solving real world problems than a substantial amount of developers. I can only assume the other AI companies are just going to copy the approach shortly too.
We built something like this by hand without much difficulty for a product concept. We'd initially used LangGraph but we ditched it and built our own out of revenge for LangGraph wasting our time with what could've simply been an ordinary python function.
Never again committing to any "framework", especially when something like Claude Code can write one for you from scratch exactly for what you want.
We have code on demand. Shallow libraries and frameworks are dead.
i noticed the same, so in the README, I describe `cord` as a protocol:
```
This repo is one implementation of the Cord protocol. The protocol itself — five primitives, dependency resolution, authority scoping, two-phase lifecycle — is independent of the backing store, transport, and agent runtime. You could implement Cord with Redis pub/sub, Postgres for multi-machine coordination, HTTP/SSE instead of stdio MCP, or non-Claude agents. See RFC.md for the full protocol specification.
```
langchain is stuck in innovator dilemma - it was built for gpt 3.5, or 4, it needs a different design for todays models, but cant evolve because of existing users and backward compatibility
just like jQuery still exists and is being actively developed
This is a vibeslop project with a vibeslop write-up.
Trees? Trees aren't expressive enough to capture all dependency structures. You either need directed acyclical graphs or general directed graphs (for iterative problems).
Based on the terminology you use, it seems you've conflated the graphs used in task scheduling with trees used in OS process management. The only reason process trees are trees are for OS-specific reasons (need for a single initializing root process, need to propagate process properties safely) . But here you're just solving a generic problem, trees are the wrong data structure.
- You have no metrics for what this can do
- No reason given for why you use trees (the text just jumps from graph to trees at one point)
- None of the concepts are explained, but it's clearly just the UNIX process model applied to task management (and you call this 60 year old idea "genuinely new"!)
DAG-scheduling is a solved issue. If you really need to impement an autonomous LLM agent workflow engine, use a DAG. Just have the LLM declare the edges.
My suggestion is to not do any of that but use a traditional workflow engine / project management tools if you need to boost productivity in processes /projects. You're code quality will improve and your programming skills won't deteriorate. At no cost in productivity.
I wonder if the “spawn” API is ever preferable over “fork”. Do we really want to remove context if we can help it? There will certainly be situations where we have to, but then what you want is good compaction for the subagent. “Clean-slate” compaction seems like it would always be suboptimal.
Is there any reason to explicitly have this binary decision.
Instead of single primitive where the parent dynamically defines the childs context. Naturally resulting in either spawn or fork or anything in between.
I built something similar that uses an event-delegation framework built around a pub/sub architecture where custom hats (personas) define specialized workflows through topic-based subscriptions.
The tasks tool is designed to validate a DAG as input, whose non-blocked tasks become cheap parallel subagent spawns using Erlang/OTP.
It works quite well. The only problem I’ve faced is getting it to break down tasks using the tool consistently. I guess it might be a matter of experimenting further with the system prompt.
Maybe, but "really smart" is doing a lot of work there. Even if the model is 10x better, you still need something to manage tool failure, retries, rate limits, cost budgets, human approval gates, and state persistence across long-running tasks. Those aren't intelligence problems, they're infrastructure problems. The framework isn't going away, it's just going to get a lot thinner.
This approach seems interesting, but in my experience, a single "agent" with proper context management is better than a complicated agent graph. Dealing with hand-off (+ hand back) and multiple levels of conversations just leaves too much room for critical information to get siloed.
If you have a narrow task that doesn't need full context, then agent delegation (putting an agent or inference behind a simple tool call) can be effective. A good example is to front your RAG with a search() tool with a simple "find the answer" agent that deals with the context and can run multiple searches if needed.
I think the PydanticAI framework has the right approach of encouraging Agent Delegation & sequential workflow first and trying to steer you away graphs[0]
Single agent with good context management beats a sloppy multi-agent setup every time, I agree. But you hit a wall around 100-120k tokens of accumulated context where the model starts losing the thread. It'll forget a constraint from early in the conversation or subtly contradict a decision it made 50 tool calls ago. That's when you need to split, not because the architecture is elegant, but because the model literally can't hold it all anymore. The question is whether you split into a pre-planned graph or let the agent decide when to split. Both have real tradeoffs.
If context window is infinite and performance isn't constrained, the subagent stuff isn't necessary. Until then, harnesses are for context management and parallelism.
Opencode getting fork was such a huge win. It's great to be able to build something out, then keep iterating by launching new forks that still have plenty of context space available, but which saw the original thing get built!
That’s what this is. It’s just defining two types of subagent relationship (spawned and forked) and providing the minimal MCP API for controlling them. It’s up to the LLMs when and how to use subagents.
They do. But with fine tuning for your use case you will get more power over what they do and if you do it right (no idea yet, if this tool here will really help) - you will get better results.
Why can’t you just give access to all tools to all subagents? That’s more general than what you’ve done. Surely it can figure out how to backtrack or keep context?
But I do like you approach and I feel this is the next step.
Trust is not objective. It's built between parties over time by looking at actions and the results of those actions. In other words, it's entirely subjective based on what's happened between the parties involved. You haven't built that trust with AI agents, or the agents have done things to lose that trust (assuming you've tried), but others have. You can't just dismiss their experience as invalid compared to your own.
Has the author (not OP) written anything on this topic themselves? This is a blunt comment, because I am fed up with being asked to read LLM content that the prompter thinks is novel and worthwhile because they don't know better.
I can forgive (even root for) someone who puts in the effort themselves to understand a problem and write about it, even if they fall short or miss. They have skin in the game. I have little patience for someone who doesn't understand the disproportionate burden generated content places on the READER.
I can certainly tell they've put the model through the ringer to be terse and use simple language, etc. But I am struggling to separate the human ideas from the vibed ones, and the tone of the whole thing is the usual LLM elevator pitch with "hushed reverence" * "movie trailer cadence".
But "spawn/fork" is just a different way of labeling the fairly-well-understood tactic (I won't call it a strategy) of just how much context to provide sub-agents. Claude Code already "spawns" everytime it does an explore. It can do this concurrently, too.
Beyond that, they seem to express wonder at how well models can use tools:
> In the example above, the agent chose spawn for the independent research tasks and fork for the analysis that needs everything. It made this choice on its own — the model understands the distinction intuitively.
Emphasis mine. They (or the model whose output they blindly published) are anthropomorphizing software that is already designed to work this way. They gave it "fork" and "spawn" tools. Are they claiming they didn't describe exactly how they were supposed to be used in the tool spec?
The criticism about the labeling is valid and I oversold. For clarity, this is what the agent sees:
`spawn`: "Create a spawned child node under your node."
`fork`: "Create a forked child node (inherits parent context) under your node."
The novelty is less about the distinction between the two, it's the tree generation. I would have served you better, if I just left out the parts that aren't critical to the novelty. Thank you for taking the time to comment.
And thanks for taking the criticism.
In all honesty, "would have written it myself but I was too eager to get it out the door" doesn't really make sense to me. You're acknowledging that you took a shortcut to get it out the door (blog post as tech debt is a new one!) - does that mean you'd like to write something up yourself eventually?
I hope so, and would like to read it. In particular, since this is presented as research, I'd be very interested to read about your experimental observations that show the risks/costs/edge cases/limits of this pattern. As it stands, there's no indication of self-critique or really any process of systematic evaluation of the solution.
It is my pleasure to write for you.
https://www.june.kim/cord-human
I enjoyed this! Thank you for taking the time to write it. I like it because I identify with your experiences way more than I do with the standard AI braggy boilerplate. Cheers.
And I deserve and accept the gentle snark at the beginning. I will be sure to let you know when I post something, I'll take your notes :)
on Wednesday, I had a dream about agents. Thursday evening, I talked to Claude about trees. That same night, I pushed out the post. There wasn't much rigor involved but yes I will explore more and report back to you!
Nice one.
You should also try to make context query the first class primitive.
Context query parameter can be natural language instruction how to compact current context passed to subagent.
When invoking you can use values like "empty" (nothing, start fresh), "summary" (summarizes), "relevant information from web designer PoV" (specific one, extract what's relevant), "bullet points about X" etc.
This way LLM can decide what's relevant, express it tersly and compaction itself will not clutter current context – it'll be handled by compaction subagent in isolation and discarded on completion.
What makes it first class is the fact that it has to be built in tool that has access to context (client itself), ie. it can't be implemented by isolated MCP because you want to avoid rendering context as input parameter during tool call, you just want short query.
Ie. you could add something like:
depends_on is also based on context query but in this case it's a map where keys are subagent conversation ids that are blockers to perform this handed over task and value is context query what to extract to inject.The actual post and this comment shows how early we are when simple and obvious ideas look novel when first conceptualizing them. Nothing against these ideas though, they are indeed good.
Thank you for the suggestion, I will explore this in the next iteration. I'm learning how to translate how humans do context management into how agents should do them
This kind of research is underrated. I have a strong feeling that these kinds of harness improvements will lead to solving whole classes of problems reliably, and matter just as much as model training.
Imposing a strict, discrete topology—like a tree or a DAG—is the only viable way to build reliable systems on top of LLMs.
If you leave agent interaction unconstrained, the probabilistic variance compounds into chaos. By encapsulating non-deterministic nodes within a rigidly defined graph structure, you regain control over the state machine. Coordination requires deterministic boundaries.
Honest question, have you actually run a DAG-based agent system past like 8 nodes? In my experience the topology isn't what gives you reliability, it's the error handling at each node. You can have the most beautiful DAG in the world and it still falls apart when node 3 returns something subtly wrong and nodes 4-8 all proceed as if everything's fine. The structure gives you the illusion of control.
The article addresses this:
> This made sense when agents were unreliable. You’d never let GPT-3 decide how to decompose a project. But current models are good at planning. They break problems into subproblems naturally. They understand dependencies. They know when a task is too big for one pass.
> So why are we still hardcoding the decomposition?
Sure, decomposition is already in the pre-training corpus. and then we can do some "instruction-tuning" on top. This is fine for the last mile, but that's it. I would consider this unaddressed and after with the root comment.
[dead]
Not necessarily - I've been having a lot of success with an event-delegation model.
https://github.com/mikeyobrien/ralph-orchestrator
My gut says you are correct though cycles can be permitted given a boundary condition.
Yep, that's what I do with https://github.com/mikeyobrien/ralph-orchestrator
This is truly dope.
I've been playing with a closely related idea of treating the context as a graph. Inspired by the KGoT paper - https://arxiv.org/abs/2504.02670
I call this "live context" because it's the living brain of my agents
Feels very AI written in a way that makes it annoying to read with all the repetitive short sentences.
Neat concept though, would be cool to see some tests of performance on some tasks.
thanks, I would've hand-written the whole thing myself but I was way too eager to get it out the door!
Not exactly a surprise Claude did this out of the box with minimal prompting considering they’ve presumably been RLing the hell out of it for agent teams: https://code.claude.com/docs/en/agent-teams
interestingly, I discovered that running `claude` sessions inside `claude` is disabled by default via env vars.
Every time i see some new orchestrator framework worth more than a few hundred loc i cringe so hard. Reddit is flooded with them on the daily and HN has them on the front page occasionally.
My current setup is this;
- `tmux-bash` / `tmux-coding-agent`
- `tmux-send` / `tmux-capture`
- `semaphore_wait`
The other tools all create lockfiles and semaphore_wait is a small inotify wrapper.
They're all you need for 3 levels of orchestration. My recent discovery was that its best to have 1 dedicated supervisor that just semaphore_wait's on the 'main' agent spawning subagents. Basically a smart Ralph-wiggum.
https://github.com/offline-ant/pi-tmux if anybody is intrested.
The tmux + lockfile approach is underrated. We went through a whole phase of building proper orchestration infra and ended up ripping most of it out. The overhead of coordinating agents through a framework is often worse than just letting them talk through the filesystem. The dirty secret of multi-agent systems is that the coordination layer is usually where the bugs live, not the agents themselves.
you cringe while simultaneously posting a github link with your “current setup” - do you see the irony?
Historically, Claude code used sequential planning with linear dependencies using tools like TodoWrite, TodoRead. There are open source MCP equivalents of TodoWrite.
I’ve found both the open source TodoWrite and building your own TodoWrite with a backing store surprisingly effective for Planning and avoiding developer defined roles and developer defined plans/workflows that the author calls in the blog for AI-SRE usecases. It also stops the agent from looping indefinitely.
Cord is a clever model and protocol for tree-like dependencies using the Spawn and Fork model for clean context and prior context respectively.
I have yet to read this article (in full), but I love trees! As an amateur AST transformation nerd. Kinda related but I’ve been trying to figure out how to generalize the lessons learned from this experiment in autogenerating massive bilingual dictionary and phrasebook datasets: https://youtu.be/nofJLw51xSk
Into a general purpose markup language + runtime for multi step LLM invocations. Although efforts so far have gotten nowhere. I have some notes on my GitHub profile readme if anyone curious: https://github.com/colbyn
Here’s a working example: https://github.com/colbyn/AgenticWorkflow
(I really dislike the ‘agentic’ term since in my mind it’s just compilers and a runtime all the way down.)
But that’s more serial procedural work, what I want is full blown recursion, in some generalized way (and without liquid templating hacks that I keep restoring to), deeply needed nested LLM invocations akin to how my dataset generation pipeline works.
PS
Also I really dislike prompt text in source code. I prefer to factor in out into standalone prompt files. Using the XML format in my case.
Claude basically does this now (including deciding when to use subagents, tools, and agent teams). I built a similar thing a month ago and saw the writing on the wall.
I agree, Claude does spawn subagents but subagents don't spawn sub-subagents.
This is the interesting constraint though. Have you found going deeper than 2 levels actually helps? Every time I've let agents spawn recursively the quality at depth 3+ drops off a cliff. Not because the model is dumber, the accumulated context drift means the deep child is solving a slightly different problem than what the root intended. Curious if your tree structure has a natural depth limit in practice or if it scales.
Are you sure? In Opencode they can, but it's hard to track them then (say, if you want to steer them) -- you have to click through. It would be nice to have a dynamic execution graph alongside the text.
[dead]
This is the comment I was looking for. In the last month or so this is how Claude Code represents tasks, as a DAG of objectives, built from plan mode.
Yeah exactly. I noticed Claude start doing exactly this a month ago too. It recursively breaks problems down while allowing you to either change direction at each level or keep going. This is where claude jumped up to be legitimately better at solving real world problems than a substantial amount of developers. I can only assume the other AI companies are just going to copy the approach shortly too.
We built something like this by hand without much difficulty for a product concept. We'd initially used LangGraph but we ditched it and built our own out of revenge for LangGraph wasting our time with what could've simply been an ordinary python function.
Never again committing to any "framework", especially when something like Claude Code can write one for you from scratch exactly for what you want.
We have code on demand. Shallow libraries and frameworks are dead.
i noticed the same, so in the README, I describe `cord` as a protocol:
``` This repo is one implementation of the Cord protocol. The protocol itself — five primitives, dependency resolution, authority scoping, two-phase lifecycle — is independent of the backing store, transport, and agent runtime. You could implement Cord with Redis pub/sub, Postgres for multi-machine coordination, HTTP/SSE instead of stdio MCP, or non-Claude agents. See RFC.md for the full protocol specification. ```
This is one of the worst takes I've ever heard.
There's a reason industries have standards. If you replace established libraries with vibecoded alternatives you will have:
- less documentation
- less tested code
- no guarantees it's doing the right thing
- a dice roll for whether it works this time on this project
- a bad time in general
langchain is stuck in innovator dilemma - it was built for gpt 3.5, or 4, it needs a different design for todays models, but cant evolve because of existing users and backward compatibility
just like jQuery still exists and is being actively developed
Or maybe it was poorly designed from the start, for there were no precedents.
This is a vibeslop project with a vibeslop write-up.
Trees? Trees aren't expressive enough to capture all dependency structures. You either need directed acyclical graphs or general directed graphs (for iterative problems).
Based on the terminology you use, it seems you've conflated the graphs used in task scheduling with trees used in OS process management. The only reason process trees are trees are for OS-specific reasons (need for a single initializing root process, need to propagate process properties safely) . But here you're just solving a generic problem, trees are the wrong data structure.
- You have no metrics for what this can do - No reason given for why you use trees (the text just jumps from graph to trees at one point) - None of the concepts are explained, but it's clearly just the UNIX process model applied to task management (and you call this 60 year old idea "genuinely new"!)
no solution is final, but if you have a better working solution, please share!
DAG-scheduling is a solved issue. If you really need to impement an autonomous LLM agent workflow engine, use a DAG. Just have the LLM declare the edges.
My suggestion is to not do any of that but use a traditional workflow engine / project management tools if you need to boost productivity in processes /projects. You're code quality will improve and your programming skills won't deteriorate. At no cost in productivity.
I wonder if the “spawn” API is ever preferable over “fork”. Do we really want to remove context if we can help it? There will certainly be situations where we have to, but then what you want is good compaction for the subagent. “Clean-slate” compaction seems like it would always be suboptimal.
This is my question also but a bit different.
Is there any reason to explicitly have this binary decision.
Instead of single primitive where the parent dynamically defines the childs context. Naturally resulting in either spawn or fork or anything in between.
That actually sounds even better than the binary. Thanks for the suggestion!
context rot and bias removal would be two good reasons to start a freshly spawned agent.
I built something similar that uses an event-delegation framework built around a pub/sub architecture where custom hats (personas) define specialized workflows through topic-based subscriptions.
https://github.com/mikeyobrien/ralph-orchestrator
My small agent harness[0] does this as well.
The tasks tool is designed to validate a DAG as input, whose non-blocked tasks become cheap parallel subagent spawns using Erlang/OTP.
It works quite well. The only problem I’ve faced is getting it to break down tasks using the tool consistently. I guess it might be a matter of experimenting further with the system prompt.
[1]: https://github.com/matteing/opal
all of these frameworks will go away once the model gets really smart. it will just be tool search, tools, and the model
in the short run, ive found the open ai agents one to be the best
Maybe, but "really smart" is doing a lot of work there. Even if the model is 10x better, you still need something to manage tool failure, retries, rate limits, cost budgets, human approval gates, and state persistence across long-running tasks. Those aren't intelligence problems, they're infrastructure problems. The framework isn't going away, it's just going to get a lot thinner.
This approach seems interesting, but in my experience, a single "agent" with proper context management is better than a complicated agent graph. Dealing with hand-off (+ hand back) and multiple levels of conversations just leaves too much room for critical information to get siloed.
If you have a narrow task that doesn't need full context, then agent delegation (putting an agent or inference behind a simple tool call) can be effective. A good example is to front your RAG with a search() tool with a simple "find the answer" agent that deals with the context and can run multiple searches if needed.
I think the PydanticAI framework has the right approach of encouraging Agent Delegation & sequential workflow first and trying to steer you away graphs[0]
[0]:https://ai.pydantic.dev/graph/
Single agent with good context management beats a sloppy multi-agent setup every time, I agree. But you hit a wall around 100-120k tokens of accumulated context where the model starts losing the thread. It'll forget a constraint from early in the conversation or subtly contradict a decision it made 50 tool calls ago. That's when you need to split, not because the architecture is elegant, but because the model literally can't hold it all anymore. The question is whether you split into a pre-planned graph or let the agent decide when to split. Both have real tradeoffs.
This isnt true for big code bases. Subagents or orchestration become vital for context handholding
yeah i think sub agents are needed, missed that in my comment
If context window is infinite and performance isn't constrained, the subagent stuff isn't necessary. Until then, harnesses are for context management and parallelism.
I don’t think so. The harness matters a lot for the task at hand, and some harnesses are much better than others for some kinds of problems.
Cool I made this thing a while back but I really like your fork spawn parallelism
https://github.com/waynenilsen/crumbler
This uses recursive task decomposition but is single thread by design. Honestly fast enough for me and makes it easier to reason about
Looks like everyone is trying to solve the same problem - here is another example I've been trying to wrap my head around lately:
Brainfile - An open protocol for agent-to-agent task coordination.
https://brainfile.md/
Well worth a look imo
Strong agree about the value of fork.
Opencode getting fork was such a huge win. It's great to be able to build something out, then keep iterating by launching new forks that still have plenty of context space available, but which saw the original thing get built!
Can't the AI just figure out by itself how and when to launch agents?
That’s what this is. It’s just defining two types of subagent relationship (spawned and forked) and providing the minimal MCP API for controlling them. It’s up to the LLMs when and how to use subagents.
They do. But with fine tuning for your use case you will get more power over what they do and if you do it right (no idea yet, if this tool here will really help) - you will get better results.
Why can’t you just give access to all tools to all subagents? That’s more general than what you’ve done. Surely it can figure out how to backtrack or keep context?
But I do like you approach and I feel this is the next step.
context rot. The human equivalent is saying, why cant the CEO write all the code and reply to all the emails
i disagree, the only different thing in your case is the SQL table that contains the tree. that's hardly 1 page. it makes no difference to context.
Whoa I didn't my blog expect to hit the front page! Hi HN!
Doesn't codex already do this when it decides whether to use subagents, and what prompt to give each subagent?
Yes, but as far as I'm aware, subagents can't spawn their own subagents, so the root agent tends to grow its context linearly
still dont see why i need any of this over the langchain / langgraph ecosystem
While those have their place, they are inherently rigid.
I love this. I always imagined more capable agent systems that have graph-like qualities.
One agent can't even be trusted to think autonomously much less a tree of them
Trust is not objective. It's built between parties over time by looking at actions and the results of those actions. In other words, it's entirely subjective based on what's happened between the parties involved. You haven't built that trust with AI agents, or the agents have done things to lose that trust (assuming you've tried), but others have. You can't just dismiss their experience as invalid compared to your own.
Agree
This is precisely how the newly released Claude agent teams work.
Not to be confused with:
cord - The #1 AI-Powered Job Search Platform for people in tech
[dead]
[dead]
[dead]
[flagged]
[flagged]
>40% of total tokens go to coordination rather than actual task completion once you get past 3-4 agents
Remarkably similar to humans.
sounds like a research project on the actual % of tokens allocated towards overhead
^ AI slop
[flagged]
[flagged]
Would those agents happened to be named frk_ai_8b2e and that platform news.ycombinator.com?
Glad you said it first. I thought it was particular comment length and then two back to back same minute comments to be strange.