Tendril – a self-extending agent that builds and registers its own tools

Apr 27, 2026•walmsles•View Original

TL;DR Highlight

Give an agent just three tools and let it build and register the rest in a growing registry.

Who Should Read

An agent developer whose agents pick wrong tools from a huge toolset or burn tokens re-solving the same problems every session.

Core Mechanics

The model always sees just three tools
- Instead of a big bag of tools, only three bootstrap tools are exposed.
- The tool surface stays fixed while capabilities grow in the registry.
If no tool exists, it builds one without asking
- It first checks the registry with searchCapabilities.
- If nothing fits, it writes a tool, registers it, and runs it by name.
- The system prompt forbids asking "would you like me to create a tool?"
The registry is just files
- It is stored as index.json plus tools/*.ts, and you can inspect, edit, or delete it.
- Each entry has name, capability, triggers, and suppression.
Direct code execution is blocked
- Per the author, the agent can only run registered capabilities.
- So every task forces it to write a tool, define its invocation conditions, and register it.
Execution runs in a Deno sandbox
- Tools run in a Deno subprocess with permission flags.
- Default timeout is 45000ms, and empty allowedDomains means unrestricted network.
Stack is Strands plus Bedrock Claude
- It uses AWS Strands Agents SDK (TypeScript), Claude on Bedrock, and a Tauri desktop shell.
- Agent and UI talk via JSON-RPC 2.0/NDJSON over stdio.

Community Reactions

Five local models all failed, per the author
- Qwen3-8B, Gemma 4, Mistral Small 3.1, Devstral Small 2, and xLAM-2 were tried and none passed.
- The author says the failure modes are written up separately.
Concern that the registry fills with noise and inconsistent APIs
- One commenter expected the registry to become noise after dozens of sessions, with overly task-specific tools.
- Another said scaling complexity is the core issue.
Similar setups people built themselves
- One company's LLM searches a wiki of Saved Programs first, reuses one if found, or generates and offers to save one.
- Another commenter creates a Claude skill or subagent for recurring problems, mentioned in one line in a global md, and says it works for a dozen or two.
Pushback on codifying WHEN as rules
- One commenter said the IF found / IF not found system prompt is still an if-then heuristic.
- They reported better generalization by describing the situation and letting the model decide.

How to Apply

Make a tool registry index.json
- Pick one recurring task for your agent and describe it with name, capability, triggers, and suppression in index.json.
- Paste the prompt below so the agent searches first and builds when nothing matches.
Cut exposed tools to a few bootstrap tools
- If tool selection is shaky, expose only search, register, and execute.
- Store the rest as files and load them via search.
Add permission limits to the execution sandbox
- Run generated code in a Deno subprocess with a timeout and allowedDomains.
- Allow only the domains each tool needs.
Set cleanup rules for the registry
- Periodically open index.json and merge or delete duplicate or overly specific tools.
- Write invocation conditions in triggers and suppression as situation descriptions.

Try It

copy & run

# Paste into Claude Code or ChatGPT (simplified version of the Tendril approach)

You are an agent that uses a tool registry. index.json in the working folder is the registry.
Fields per entry: name (snake_case), capability (one sentence), triggers, suppression.

Rules:
1. On a request, first search index.json for a relevant tool.
2. If found, run tools/{name}.ts and answer.
3. If not found, do not ask "shall I create one?"; write the tool, save it in tools/, register it in index.json, then run it.
4. If it fails, read the error, fix the code, and retry.
5. Do not answer from training data when a tool could get live information.

Request: {{request}}

Terminology

Agent Capability patternA pattern where the model discovers, builds, and reuses tools across sessions. Think of a function library that grows with use.

capability registryA list of tool names, descriptions, and invocation conditions. It is the agent's tool catalog.

Bootstrap toolsThe minimal tools given from the start. They are used to build all the others.

NDJSONNewline-delimited JSON messages, one per line. Like a log format where each line is a message.

Deno sandboxA Deno subprocess that runs code with permission flags limiting file and network access. A isolated room for execution.

Related Resources

Tendril – a self-extending agent that builds and registers its own tools

Apr 27, 2026•walmsles•View Original

TL;DR Highlight

Give an agent just three tools and let it build and register the rest in a growing registry.

Who Should Read

An agent developer whose agents pick wrong tools from a huge toolset or burn tokens re-solving the same problems every session.

Core Mechanics

The model always sees just three tools
- Instead of a big bag of tools, only three bootstrap tools are exposed.
- The tool surface stays fixed while capabilities grow in the registry.
If no tool exists, it builds one without asking
- It first checks the registry with searchCapabilities.
- If nothing fits, it writes a tool, registers it, and runs it by name.
- The system prompt forbids asking "would you like me to create a tool?"
The registry is just files
- It is stored as index.json plus tools/*.ts, and you can inspect, edit, or delete it.
- Each entry has name, capability, triggers, and suppression.
Direct code execution is blocked
- Per the author, the agent can only run registered capabilities.
- So every task forces it to write a tool, define its invocation conditions, and register it.
Execution runs in a Deno sandbox
- Tools run in a Deno subprocess with permission flags.
- Default timeout is 45000ms, and empty allowedDomains means unrestricted network.
Stack is Strands plus Bedrock Claude
- It uses AWS Strands Agents SDK (TypeScript), Claude on Bedrock, and a Tauri desktop shell.
- Agent and UI talk via JSON-RPC 2.0/NDJSON over stdio.

Community Reactions

Five local models all failed, per the author
- Qwen3-8B, Gemma 4, Mistral Small 3.1, Devstral Small 2, and xLAM-2 were tried and none passed.
- The author says the failure modes are written up separately.
Concern that the registry fills with noise and inconsistent APIs
- One commenter expected the registry to become noise after dozens of sessions, with overly task-specific tools.
- Another said scaling complexity is the core issue.
Similar setups people built themselves
- One company's LLM searches a wiki of Saved Programs first, reuses one if found, or generates and offers to save one.
- Another commenter creates a Claude skill or subagent for recurring problems, mentioned in one line in a global md, and says it works for a dozen or two.
Pushback on codifying WHEN as rules
- One commenter said the IF found / IF not found system prompt is still an if-then heuristic.
- They reported better generalization by describing the situation and letting the model decide.

How to Apply

Make a tool registry index.json
- Pick one recurring task for your agent and describe it with name, capability, triggers, and suppression in index.json.
- Paste the prompt below so the agent searches first and builds when nothing matches.
Cut exposed tools to a few bootstrap tools
- If tool selection is shaky, expose only search, register, and execute.
- Store the rest as files and load them via search.
Add permission limits to the execution sandbox
- Run generated code in a Deno subprocess with a timeout and allowedDomains.
- Allow only the domains each tool needs.
Set cleanup rules for the registry
- Periodically open index.json and merge or delete duplicate or overly specific tools.
- Write invocation conditions in triggers and suppression as situation descriptions.

Try It

copy & run

# Paste into Claude Code or ChatGPT (simplified version of the Tendril approach)

You are an agent that uses a tool registry. index.json in the working folder is the registry.
Fields per entry: name (snake_case), capability (one sentence), triggers, suppression.

Rules:
1. On a request, first search index.json for a relevant tool.
2. If found, run tools/{name}.ts and answer.
3. If not found, do not ask "shall I create one?"; write the tool, save it in tools/, register it in index.json, then run it.
4. If it fails, read the error, fix the code, and retry.
5. Do not answer from training data when a tool could get live information.

Request: {{request}}

Terminology

Agent Capability patternA pattern where the model discovers, builds, and reuses tools across sessions. Think of a function library that grows with use.

capability registryA list of tool names, descriptions, and invocation conditions. It is the agent's tool catalog.

Bootstrap toolsThe minimal tools given from the start. They are used to build all the others.

NDJSONNewline-delimited JSON messages, one per line. Like a log format where each line is a message.

Deno sandboxA Deno subprocess that runs code with permission flags limiting file and network access. A isolated room for execution.

Tendril – a self-extending agent that builds and registers its own tools

TL;DR Highlight

Who Should Read

Core Mechanics

Community Reactions

How to Apply

Try It

Terminology

Related Papers

Related Resources

Tendril – a self-extending agent that builds and registers its own tools

TL;DR Highlight

Who Should Read

Core Mechanics

Community Reactions

How to Apply

Try It

Terminology

Related Papers

Related Resources