mission.md

Mission - Agent Workbench: Why Capable Models Still Fail

Goal

Run the same small repo task twice, once prompt-only and once with the seven workbench surfaces wired in, and emit a failure-mode report that maps each missed surface to the symptom it caused.

Inputs

• A stub agent and a tiny FastAPI-style handler to validate
• The seven-surface list (instructions, state, scope, feedback, verification, review, handoff)

Deliverables

• code/main.py that runs both pipelines back to back
• failure_modes.json summarizing the prompt-only run
• One-line verdict for the workbench run

Acceptance

• python3 code/main.py exits zero
• Output shows a side-by-side log of the two runs
• failure_modes.json lists every missed surface with the matching symptom

Out of scope

• Calling a real model. The stub is rule-based on purpose.
• Building any one surface in depth. That is what the next eleven lessons are for.

References

• docs/en.md - full lesson
• code/main.py - reference implementation
• outputs/skill-workbench-audit.md - extracted skill