Tendril, a self‑extending sandbox, lets a model discover, build, and reuse tools without human prompts. Using AWS Strands Agents SDK and Tauri, it checks a registry, writes missing capabilities, and executes them automatically. The result is a growing toolset that adapts each session.

Built on a Bedrock Claude model, Tendril provides only three bootstrap tools, forcing the agent to search a registry rather than choose from a large menu. When a request arrives, the system first looks for an existing tool, loads it if found, or compiles a new one in TypeScript and runs it in a Deno sandbox. The registry lives as JSON files in a workspace, making inspection and editing trivial.

The architecture layers a Rust‑based Tauri shell, a React UI, and a Node.js agent binary. Communication follows JSON‑RPC over NDJSON, keeping the agent a lightweight sidecar. Developers can configure AWS credentials, model IDs, and sandbox limits via a single JSON file.

Tendril demonstrates the Agent Capability pattern in practice, showing how autonomous agents can evolve their toolset over time while keeping the surface minimal and under developer control.