What problem does it solve? Autonomous agents running long research or experimentation loops can corrupt their objectives, game their own metrics, lose state after context compaction, or take irreversible actions without approval. This Skill provides the design patterns for building a harness around an agent so it can run safely for extended periods. ## Core Features & Use Cases - Surface Classification: Define locked, editable, append-only, and human-controlled surfaces so agents cannot modify the evaluators that score them. - Durable State & Logging: Externalize plans, results, and rejected attempts into thread logs and TSV/JSONL files that survive context compaction. - Search Discipline & Governance: Add novelty gates, pruning rounds, rollback policies, and explicit human approval boundaries for merges and deploys. - Use Case: You are building a background agent that reads new research posts and drafts skill updates as pull requests. Use this Skill to lock the evaluation rubrics, set up a results log, and ensure the agent prepares PRs but never merges them without human review. ## Quick Start Design a harness for my autonomous research agent with locked evaluation rubrics, an append-only results log, and human approval required before any PR is merged.
