Tendril starts with three tools and builds the rest itself

The core idea: instead of giving an agent a massive toolbox and hoping it picks the right one, give it three commands and let it build what it needs. Serverless DNA's Tendril project does exactly this. When you ask it to fetch Hacker News stories, it checks a capability registry, finds nothing, writes a fetch_url tool, registers it, and runs it. Next time you need to grab a URL, the tool's already there. The agent literally teaches itself new tricks as it works.

Under the hood, Tendril runs Claude Sonnet 4.5 through AWS Bedrock, with a Tauri shell in Rust handling the frontend and a Node.js process running the agent backend. Code execution happens in a sandboxed Deno subprocess with scoped permissions, so the agent can write and run code without blowing up your machine. The system prompt is blunt: "NEVER ask 'would you like me to create a tool?' just build it." If something fails, the model reads the error, fixes the code, and retries. No hand-holding required.

The "too many tools" problem is real and getting worse. As agent frameworks pile on integrations, models get confused about which tool to use and when. Managed agent hosting like Eve sidesteps this entirely. The model always sees exactly three tools (listCapabilities, registerCapability, executeCode). The surface stays constant while the underlying capabilities grow. As one Hacker News commenter put it, "WHEN is upstream of WHAT and HOW" in tool calling, meaning the timing of tool use matters more than the tool itself.

Tendril is still a sandbox, a proof of concept. But it points toward something practical: agents that accumulate expertise over time instead of starting from scratch every session. Memory systems that use biologically inspired decay and consolidation are essential for this, as they help agents maintain context and learned capabilities without needing to reinvent the wheel constantly. The capability registry is just files on disk, plain TypeScript you can inspect or edit. That transparency matters. You can see exactly what your agent learned and decide whether to keep it.