Skip to content

Agent Tools

Tools turn model output into action. Each tool should state accepted inputs, what it may modify, how it fails, and how it can be tested alone.

Tool typeDefault boundary
Read-only lookupUsually safe to open first.
File writeLimit directories and preserve diffs.
Shell commandDangerous commands need confirmation.
External APINeeds timeout, budget, and logs.
Inspect the repository and explain the plan first. Do not edit files. Stop before writes or destructive commands.

A tool is not just a helper. It is a capability with inputs, outputs, side effects, and a failure mode. Test the tool outside the agent when possible, then let the agent call it.

Tool canBoundary to add
Read filesLimit the working directory.
Write filesReview diffs before commit.
Run commandsConfirm destructive commands.
Call APIsUse timeouts, budgets, and logs.
Publish or sendKeep human approval in the path.
Inspect the repository and propose a change. Do not edit files yet. Stop before running any destructive command.