> The registry grows with use. Every session is smarter than the last.
This feels a bit like one of those “now you have two problems” solutions. After a few dozen sessions I would expect the tool registry to be full of “noise” for most prompts. I would also expect most tools to be extremely specific to the task at hand, leading to redundancy and ultimately poor programmability due to inconsistencies between tool APIs.
A while back I realized OpenClaw was Claude Code in a trenchcoat, except that Claude Code is pretty good at extending itself without breaking itself. (Note: haven't used OC since February, maybe it's solid and reliable now.)
Of course, being reliable and reliably extensible is the whole point, which means Claude Code made a better OC than OC did! I found this very amusing for some reason.
Also you can put it (or your agent of choice, e.g. codex works too) in a Telegram bot in like 50 lines of code which is a lot of fun.
> It solves the problem the originating user asked it to
Interesting. And is there a mechanism to go back and "fix" the tools after they are published? What happens if the tool decided to use the "id" attribute to click on buttons and now you have a new website that follows a different pattern to find the right target?
I agree that "correctness" of a tool could have different meaning depending on the context of the problem though (e.g. would you consider OOM a correctness bug even if it addresses the user's ask?)
Everyone is just taking a round about way to get there. The workflow/program as "tools" approach is the right one. Agents skills are more or less in that same direction.
I like that you approach the question of "when" in regards to tool calls. I've become frustrated that most agent frameworks don't acknowledge it in their design philosophy.
WHEN is upstream of WHAT and HOW. You can have perfect tool descriptions and perfect call signatures, but if the model can't read the situation to know whether the moment calls for any tool at all, you get either over-firing (agent burns tokens trying to "help") or under-firing (agent waits to be addressed and acts like a chatbot, not an autonomous participant).
I have had a lot of success when I refrain from codifying WHEN as rules. "If X then fire tool Y" is a dumb heuristic with extra steps. Describe the conditions of the moment. What's been tried, what's converged, what state the work is in. Then let the model decide whether to act and which tool fits.
Rules get stale. Situation-reads generalize.
Reading the Tendril README, looks like the registration mechanic is solving a slightly different problem (the "too many tools" / context-bloat problem) by giving the agent three bootstrap tools and a growing registry. The WHEN itself still seems to be codified as rules in the system prompt ("BEFORE acting, call searchCapabilities; IF found, load and execute; IF NOT found, build yourself"). That's exactly the IF-X-THEN-Y pattern your framing seems to want to move past.
Curious whether you see the registry itself as the structured WHEN, or whether the rule-based system prompt is a starting point you intend to evolve toward something more situational.
Get outta my swamp! Just kidding, it’s cool to see other people working on this stuff.
I think right now this is still a bit too fresh out of Claude Code to be usable by anybody but the people developing it. I got to around the same point with my first tempt at building a tool registry (https://github.com/accretional/collector) and then realized I basically needed to start over with much more investment in supporting infrastructure to build the thing I really wanted.
I can go as far into the weeds as anybody would ever care to hear about this, but for the sake of brevity I’ll just say this: reflection and type systems over the network are pretty much the only way to get this stuff to work properly (I mean you could just go full MCP/Skills but then all you really have are giant blobs of markdown and unconstrained json that make integration/discovery/usability a nightmare, and require an agent in the loop to drive/integrate the tools when you really just need to give them the actual APIs and documentation). That ends up getting rather hairy, we ended up actually building a declarative meta-lexer/parser/transpiler (meta basically just meaning it’s generalized across languages and self-hosting/bootstrapped) recently (https://github.com/accretional/gluon) because it turns out building a cross-language distributed type system is rather difficult. But reflection alone gets you halfway there as far as benefits.
We built something similar[1], including integrated memory for debugging. It is very useful to have repeatable artifacts left behind every time you use an agent to accomplish a task.
The main design decision we took was to integrate with your existing agent instead of building a new one. Your harness, swamp, and you're off.
As an aside, building software for agents is incredibly fun.
I built this while working on a coding agent that kept starting cold every session. The deeper problem was that agent frameworks give you what a tool does and how to call it, but no structured answer to when — when should a tool fire autonomously, and when should it stay silent. That judgement is always implicit, scattered across system prompts and tool descriptions.
Tendril is a reference implementation of what I'm calling the Agent Capability pattern. It starts with three bootstrap tools and builds everything else itself. The key constraint: there's no direct code execution. The agent can only run registered capabilities, so every task forces it to write a tool, define its invocation conditions, and register it for future sessions. The registry accumulates across sessions.
I also ran the self-extending loop against five local models — Qwen3-8B, Gemma 4, Mistral Small 3.1, Devstral Small 2, Salesforce xLAM-2. None passed.
I did something that sounds similar for my home assistant.
The agent never executes anything. It has like four tools… search, request execute, request build, request update.
The tool service runs vector search against the tools catalog.
The build generalizes the requested function and runs authoring with review steps, declaring needed credentials and network access.
The adversarial reviewer can reject back to the authoring three times.
After passing, the tool is registered and embeddings are done for search. It’s live for future use.
Credentials are stored encrypted, and only get injected by the tools catalog service during tool execution. The network resources are declared so tool function execution can be better sandboxed (it’s not, yet).
The agent never has access to credentials and cannot do anything without going through vetted functions in the tool service.
Agent, author process, reviewer, embedding… all can be different models running local or remote.
Event bus, agent, tool service… all separate containers.
You can list the uses of the available tools in the AGENTS. I keep my agents on a tight leash, and self-extension runs counter to this. I would not my agent to spontaneously develop the ability to tap my bank account, for example.
I use Swival’s /learn command at the end of a session to make it write down what it got wrong, how it fixed the issue, and what it should remember next time. Works pretty well.
It can update those notes automatically, but I’ve found that even with regular nudges, models are still somewhat reluctant to do it.
So manually running /learn every now and then, especially when I can tell it didn’t take the most direct path, helps.
> The registry grows with use. Every session is smarter than the last.
This feels a bit like one of those “now you have two problems” solutions. After a few dozen sessions I would expect the tool registry to be full of “noise” for most prompts. I would also expect most tools to be extremely specific to the task at hand, leading to redundancy and ultimately poor programmability due to inconsistencies between tool APIs.
A while back I realized OpenClaw was Claude Code in a trenchcoat, except that Claude Code is pretty good at extending itself without breaking itself. (Note: haven't used OC since February, maybe it's solid and reliable now.)
Of course, being reliable and reliably extensible is the whole point, which means Claude Code made a better OC than OC did! I found this very amusing for some reason.
Also you can put it (or your agent of choice, e.g. codex works too) in a Telegram bot in like 50 lines of code which is a lot of fun.
https://github.com/a-n-d-a-i/ULTRON/blob/main/src/index.ts
Though this might get you banned from Anthropic, they haven't quite clarified that yet. (Ostensibly it defaults to extra usage now, but who knows.)
https://news.ycombinator.com/item?id=47852834
It's really cool to see that other people run into the same issues and arrive at the same conclusions/solution.
At $DAYJOB, we have an LLM-based tool and this issue of "how do we avoid burning tokens solving the same problems over again" was an early obstacle
We wound up building a very similar thing to what you call "tools" (we named them "Saved Programs").
There's a wiki the LLM searches before solving a problem, that links saved programs for past actions to their content entry.
If it finds one, it'll re-use it, otherwise it'll generate a program and offer to save it, if you think it'll be common enough.
> how do we avoid burning tokens solving the same problems over again
Letting the LLM write half baked tools is the recipe for burning more tokens.
> There's a wiki the LLM searches before solving a problem, that links saved programs for past actions to their content entry.
What's the criteria for marking an LLM written tool as useful/correct before publishing it?
> It solves the problem the originating user asked it to
Interesting. And is there a mechanism to go back and "fix" the tools after they are published? What happens if the tool decided to use the "id" attribute to click on buttons and now you have a new website that follows a different pattern to find the right target?
I agree that "correctness" of a tool could have different meaning depending on the context of the problem though (e.g. would you consider OOM a correctness bug even if it addresses the user's ask?)
it's called workflow automation: https://blog.codesolvent.com/2025/12/workflow-automation-let...
Everyone is just taking a round about way to get there. The workflow/program as "tools" approach is the right one. Agents skills are more or less in that same direction.
I like that you approach the question of "when" in regards to tool calls. I've become frustrated that most agent frameworks don't acknowledge it in their design philosophy.
WHEN is upstream of WHAT and HOW. You can have perfect tool descriptions and perfect call signatures, but if the model can't read the situation to know whether the moment calls for any tool at all, you get either over-firing (agent burns tokens trying to "help") or under-firing (agent waits to be addressed and acts like a chatbot, not an autonomous participant).
I have had a lot of success when I refrain from codifying WHEN as rules. "If X then fire tool Y" is a dumb heuristic with extra steps. Describe the conditions of the moment. What's been tried, what's converged, what state the work is in. Then let the model decide whether to act and which tool fits.
Rules get stale. Situation-reads generalize.
Reading the Tendril README, looks like the registration mechanic is solving a slightly different problem (the "too many tools" / context-bloat problem) by giving the agent three bootstrap tools and a growing registry. The WHEN itself still seems to be codified as rules in the system prompt ("BEFORE acting, call searchCapabilities; IF found, load and execute; IF NOT found, build yourself"). That's exactly the IF-X-THEN-Y pattern your framing seems to want to move past.
Curious whether you see the registry itself as the structured WHEN, or whether the rule-based system prompt is a starting point you intend to evolve toward something more situational.
Get outta my swamp! Just kidding, it’s cool to see other people working on this stuff.
I think right now this is still a bit too fresh out of Claude Code to be usable by anybody but the people developing it. I got to around the same point with my first tempt at building a tool registry (https://github.com/accretional/collector) and then realized I basically needed to start over with much more investment in supporting infrastructure to build the thing I really wanted.
I can go as far into the weeds as anybody would ever care to hear about this, but for the sake of brevity I’ll just say this: reflection and type systems over the network are pretty much the only way to get this stuff to work properly (I mean you could just go full MCP/Skills but then all you really have are giant blobs of markdown and unconstrained json that make integration/discovery/usability a nightmare, and require an agent in the loop to drive/integrate the tools when you really just need to give them the actual APIs and documentation). That ends up getting rather hairy, we ended up actually building a declarative meta-lexer/parser/transpiler (meta basically just meaning it’s generalized across languages and self-hosting/bootstrapped) recently (https://github.com/accretional/gluon) because it turns out building a cross-language distributed type system is rather difficult. But reflection alone gets you halfway there as far as benefits.
We built something similar[1], including integrated memory for debugging. It is very useful to have repeatable artifacts left behind every time you use an agent to accomplish a task.
The main design decision we took was to integrate with your existing agent instead of building a new one. Your harness, swamp, and you're off.
As an aside, building software for agents is incredibly fun.
1: https://swamp.club
I built this while working on a coding agent that kept starting cold every session. The deeper problem was that agent frameworks give you what a tool does and how to call it, but no structured answer to when — when should a tool fire autonomously, and when should it stay silent. That judgement is always implicit, scattered across system prompts and tool descriptions.
Tendril is a reference implementation of what I'm calling the Agent Capability pattern. It starts with three bootstrap tools and builds everything else itself. The key constraint: there's no direct code execution. The agent can only run registered capabilities, so every task forces it to write a tool, define its invocation conditions, and register it for future sessions. The registry accumulates across sessions.
I also ran the self-extending loop against five local models — Qwen3-8B, Gemma 4, Mistral Small 3.1, Devstral Small 2, Salesforce xLAM-2. None passed.
The failure modes were distinct enough to be worth writing up separately: https://serverlessdna.com/strands/ai-agents/agents-know-what...
Stack: AWS Strands TypeScript SDK, Bedrock (Claude Sonnet), Deno sandbox, Tauri + React desktop shell.
I did something that sounds similar for my home assistant.
The agent never executes anything. It has like four tools… search, request execute, request build, request update.
The tool service runs vector search against the tools catalog.
The build generalizes the requested function and runs authoring with review steps, declaring needed credentials and network access.
The adversarial reviewer can reject back to the authoring three times.
After passing, the tool is registered and embeddings are done for search. It’s live for future use.
Credentials are stored encrypted, and only get injected by the tools catalog service during tool execution. The network resources are declared so tool function execution can be better sandboxed (it’s not, yet).
The agent never has access to credentials and cannot do anything without going through vetted functions in the tool service.
Agent, author process, reviewer, embedding… all can be different models running local or remote.
Event bus, agent, tool service… all separate containers.
I have an url if you want to read a bit about what I did: https://dcd.fyi/agent
It’s really just meant for me, but if you’re interested in more details on anything let me know. There’s nothing super special in it.
You can list the uses of the available tools in the AGENTS. I keep my agents on a tight leash, and self-extension runs counter to this. I would not my agent to spontaneously develop the ability to tap my bank account, for example.
So basically you've built a mechanism for a model to de-compress itself.
Make a tool to fix asci diagram alignment
I use Swival’s /learn command at the end of a session to make it write down what it got wrong, how it fixed the issue, and what it should remember next time. Works pretty well.
It can update those notes automatically, but I’ve found that even with regular nudges, models are still somewhat reluctant to do it.
So manually running /learn every now and then, especially when I can tell it didn’t take the most direct path, helps.