Tendril – a self-extending agent that builds and registers its own tools
TL;DR Highlight
Tendril demonstrates a self-extending AI agent pattern by dynamically writing and registering tools when needed, creating a growing repository of capabilities with each session.
Who Should Read
Backend/full-stack developers building LLM-powered agents who face repetitive problem-solving or token waste across sessions. AI application developers designing how and when agents should utilize tools.
Core Mechanics
- Tendril serves as a reference implementation of the 'Agent Capability Pattern', where agents search existing tools, write and register new ones if missing, and execute them directly if available.
- A core design constraint is prohibiting direct code execution; agents must register tools before running them, automatically building a reusable toolset that accumulates across sessions.
- The agent loop relies on only three bootstrap tools—searchCapabilities, registerCapability, and execute—with the agent creating the rest. It's powered by AWS Strands TypeScript SDK and Bedrock (Claude Sonnet).
- Tool execution is sandboxed using Deno subprocesses, and the desktop shell is built with Tauri + React.
- Testing the self-extending loop with five local open-source models (Qwen3-8B, Gemma 4, Mistral Small 3.1, Devstral Small 2, Salesforce xLAM-2) resulted in complete failure, with each model exhibiting unique failure patterns detailed in a separate post.
- Tendril addresses the 'WHEN problem'—most agent frameworks define WHAT tools do and HOW to call them, but lack a structured approach to determine WHEN to use them. Tendril encodes this decision logic as rules within the system prompt.
- The registry structure is simple CRUD-based on index.json, and the source code is organized into agent.ts, loop/(tools, prompt, registry, sandbox), and transport/(protocol, stream, errors).
Evidence
- "Concerns were raised about the registry becoming noisy with accumulated sessions, leading to overspecialized tools, duplication, and API inconsistencies. One developer shared experience building a similar system called 'Saved Programs' to avoid token waste from repetitive problem-solving. Another developer highlighted the need for a network-based reflection and type system for effective tool registries, having built a custom distributed type system (gluon) after struggling with MCP/Skills. Criticism arose regarding the rule-based (if X then Y) approach to the 'WHEN' problem, with some finding success by describing the current state instead of prescribing rules. The importance of complexity management as the registry grows was emphasized, specifically addressing context understanding, knowledge retention, and performance monitoring."
How to Apply
- "If your agent repeatedly generates the same API calls or data processing code, introducing Tendril’s searchCapabilities → registerCapability → execute pattern can enable reuse without rebuilding. If you already use AWS Strands TypeScript SDK and Bedrock (Claude Sonnet), you can adapt Tendril’s tendril-agent/src/ structure, including the capability registry (index.json CRUD) and Deno sandbox execution layer. When building self-extending agents with local open-source models (Qwen3-8B, Gemma 4, etc.), consider that Tendril’s tests failed with all five models, suggesting a more powerful model like Claude Sonnet is currently necessary. To carry learnings across sessions, implement a routine to explicitly document what the agent learned and how it improved, similar to the '/learn command' mentioned in the comments."
Code Example
// Tendril agent loop example (based on README)
// First request: No tools → Create tool and execute
You: "fetch the top stories from Hacker News"
Tendril:
→ searchCapabilities("fetch url hacker news") // Search registry → Not found
→ registerCapability(fetch_url, code) // Write and register tool code
→ execute("fetch_url", {url: "https://..."}) // Execute registered tool
→ "Here are the top stories: ..."
// Second request: Reuse tool
You: "now fetch Lobsters and compare"
Tendril:
→ listCapabilities() // Search registry → Found fetch_url!
→ execute("fetch_url", {url: "https://lobste.rs"}) // Execute directly without rebuilding
// Directory structure
// tendril-agent/src/
// ├── agent.ts ← Strands agent + Bedrock model configuration
// ├── index.ts ← Orchestrator
// ├── loop/
// │ ├── tools.ts ← 3 bootstrap tools
// │ ├── prompt.ts ← System prompt (rules for autonomous behavior)
// │ ├── registry.ts ← Capability registry (index.json CRUD)
// │ └── sandbox.ts ← Deno subprocess sandbox execution
// └── transport/
// ├── protocol.ts ← ACP JSON-RPC over stdio
// ├── stream.ts ← SDK events → loop stage transformation
// └── errors.ts ← Provider error classificationTerminology
Related Papers
Show HN: adamsreview – better multi-agent PR reviews for Claude Code
Claude Code에서 최대 7개의 병렬 서브 에이전트가 각각 다른 관점으로 PR을 리뷰하고, 자동 수정까지 해주는 오픈소스 플러그인이다. 기존 /review나 CodeRabbit보다 실제 버그를 더 많이 잡는다고 주장하지만 커뮤니티에서는 복잡도와 실효성에 대한 회의론도 나왔다.
How Fast Does Claude, Acting as a User Space IP Stack, Respond to Pings?
Claude Code에게 IP 패킷을 직접 파싱하고 ICMP echo reply를 구성하도록 시켜서 실제로 ping에 응답하게 만든 실험으로, 'Markdown이 곧 코드이고 LLM이 프로세서'라는 아이디어를 네트워크 스택 수준까지 밀어붙인 재미있는 사례다.
Show HN: Git for AI Agents
AI 코딩 에이전트(Claude Code 등)가 수행한 모든 툴 호출을 자동으로 추적하고, 어떤 프롬프트가 어느 코드 줄을 작성했는지 blame까지 가능한 버전 관리 도구다.
Principles for agent-native CLIs
AI 에이전트가 CLI 도구를 더 잘 사용할 수 있도록 설계하는 원칙들을 정리한 글로, 에이전트가 CLI를 도구로 활용하는 빈도가 높아지면서 이 설계 방식이 실용적으로 중요해지고 있다.
Agent-harness-kit scaffolding for multi-agent workflows (MCP, provider-agnostic)
여러 AI 에이전트가 서로 역할을 나눠 협업할 수 있도록 조율하는 scaffolding 도구로, Vite처럼 설정 없이 빠르게 멀티 에이전트 파이프라인을 구성할 수 있다.
Show HN: Tilde.run – Agent sandbox with a transactional, versioned filesystem
AI 에이전트가 실제 프로덕션 데이터를 건드려도 롤백할 수 있는 격리된 샌드박스 환경을 제공하는 도구로, GitHub/S3/Google Drive를 하나의 버전 관리 파일시스템으로 묶어준다.