Guard
Something that stops a kind of failure before it reaches the end of an unattended run.
A guard is a check, a rule or a way of working that makes a failure kind impossible or visible. The guards in this map are short checked segments, read-only tests, holdout scenarios, a stop option, a near-perfect verifier, asking when underspecified, fresh starts, descriptive tool names, and leaving out repository overviews.
Kinds of this
- Ask when underspecified: Ask a person only for what only a person holds.
- Descriptive tool names: Name and describe each tool for its purpose.
- Fresh starts: Restart the agent when a condition is met, not on a calendar.
- Holdout scenarios: Checks stored outside the codebase, where the agent cannot see them.
- Leave out repository overviews: Do not add repository overviews to the agent's context.
- Near-perfect verifier: A check that is close to right, so the agent cannot solve the wrong problem.
- Read-only tests: Tests the agent can run but cannot edit.
- Short checked segments: Cut a long task into short segments and check each one.
- Stop option: A way for the agent to flag a task as impossible and hand it to a person.