What problem does it solve? Agents often drift from the behavior their instructions promise, and manually writing test cases to catch that drift is slow and incomplete. This Skill autonomously derives probes from an agent's INSTRUCTIONS and its real recorded usage, runs them against the live container, judges the responses, edits the agent file, and re-probes until behavior matches intent β no user-supplied test cases needed. ## Core Features & Use Cases - Autonomous probe generation: Mines real session and eval data from the database, then derives 8β12 probes per agent covering golden paths, edge cases, tool selection, and adversarial inputs like prompt injection. - Closed-loop hardening: Runs probes via the live HTTP API, judges PASS/FAIL by root cause, applies one targeted edit per iteration (instructions, tools, model, or history), restarts the container, and re-probes β capped at 5 iterations. - Safe state management: Snapshots and sweeps durable rows that probes leave in learning components and the platform-builder, so testing never pollutes shared stores. - Use Case: A support agent keeps issuing refunds its policy forbids. Run this Skill to generate adversarial refund probes, identify the missing instruction rule, tighten the agent file, and verify the fix with regression probes. ## Quick Start Ask the agent to run /improve-agent on a target agent slug to harden it against its stated instructions.
