12 hours ago
- The idea started with giving Pi a single tool called fabric_exec that accepts TypeScript code, allowing the model to write small programs against a typed API.
- The model's programs return only the necessary data, reducing context usage and enabling real control flow like loops and parallel execution.
- The system evolved to handle schema errors, state management, deterministic compaction, memory, and cross-session recall, with the transcript as the source of truth.
- Fabric was used to build itself: 337 commits in 12 days, 49,569 source lines and 32,059 test lines, most written by the model through the tool.
- Real-world tests included browser automation with Chrome DevTools Protocol and a 3D printing project that lasted eight days, successfully handling context compaction and failure recovery.
- The runtime cost about $1,500–$2,000 in equivalent API usage for development, with 11,155 programs executed and 93.4 million input tokens consumed.
- The key insight: giving the model a programming language inside the agent harness works better than a catalog of tiny verb tools, with strict security via sandboxed execution.