Tag: tool-use

The tool layer, not the prompt, is where an agent behaves or misbehaves. Lessons from building Zoe's tools: uniform pagination, broaden-on-empty hints, read/write parity, and treating the system prompt as policy.

Notes on turning Zoe from a copilot that talks into one that edits the document you are looking at — the bridge, reviewable edits, and the bugs that came from a model acting on stale context.