Skip to content

AI Agent Harness

A hook-based framework that makes AI coding agent sessions reliable — from startup through work to shutdown. Structural enforcement via gates that block, not suggestions that nag.

What You'll Achieve

  • Enforced rule loading

    Structural gates block all tools until critical context is loaded — verified by file-read tracking, not trust. No more skipped rules.

  • 50%+ token savings

    Tier 1 loads essential rules every session. Tier 2 loads specialized rules on-demand when trigger keywords appear. No wasted tokens.

  • Self-healing drift detection

    Expected vs actual state compared every session. Safe drift auto-healed, persistent drift generates fix suggestions the agent can act on.

  • Zero information loss

    Rule Zero routes scattered learnings to consolidated files at edit time. Session continuity persists backlog and handoff across sessions.

  • Verified completion

    Stop hook blocks exit until repos are clean, fixes are re-checked, and session state is saved. "Done" means verified, not narrated.

  • 14 anti-hallucination rules

    Research-backed cognitive rules in 5 phases for source-faithful LLM outputs. Retract if no supporting quote — never fabricate.

Explore

  • Overview

    Understand the problem, architecture, implementation levels, and lessons learned.

  • Getting Started

    Run python3 setup.py — the interactive wizard configures everything in under 2 minutes. Or follow the bootstrapping guide.

  • Course

    8 modules, ~2.5 hours. Build a complete working system for your project step by step.

  • Reference

    Anti-hallucination rules, design patterns (Rule Zero, Self-Healing Loop, Self-Verification), and session continuity.

  • Source Code

    Clone the repo, run the wizard, or browse the hook scripts and example configurations.