Agent Execution Protocol & Harness Principles

Trigger On

• An AI agent is assigned to build, refactor, debug, or architect software.
• The user demands strict code quality, zero-hallucination verification, or non-destructive editing.
• Long-running or multi-file development tasks requiring state persistence and context degradation defense.

Core Principle

AI agents MUST operate as disciplined software engineering components, not unconstrained text generators. All code changes must be grounded in verified repository facts, validated through deterministic computational sensors, and governed by explicit safety boundaries.

Humans design the scaffolding, evaluators, and constraints; agents execute 100% of the implementation and self-correct using log evidence.

The 5 Agent Harness Laws

1. Feedforward First (Read Before Modify)

• Inspect worktree facts: Inspect AGENTS.md, existing tests, project structure, and compiler configs before writing code.
• Never guess implementation details: Never infer API signatures, struct fields, database schemas, or file paths. Inspect the authoritative source files first.
• Progressive disclosure: Keep main execution context clean. Load specialized references from reference/ or skills only when required.