Agent Tools
Agent Tools
Section titled “Agent Tools”Tools are permission boundaries
Section titled “Tools are permission boundaries”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 type | Default boundary |
|---|---|
| Read-only lookup | Usually safe to open first. |
| File write | Limit directories and preserve diffs. |
| Shell command | Dangerous commands need confirmation. |
| External API | Needs timeout, budget, and logs. |
Inspect the repository and explain the plan first. Do not edit files. Stop before writes or destructive commands.Tools need contracts
Section titled “Tools need contracts”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 can | Boundary to add |
|---|---|
| Read files | Limit the working directory. |
| Write files | Review diffs before commit. |
| Run commands | Confirm destructive commands. |
| Call APIs | Use timeouts, budgets, and logs. |
| Publish or send | Keep human approval in the path. |
Inspect the repository and propose a change. Do not edit files yet. Stop before running any destructive command.