Antigravity Harness: open-source fix for coding agents that cheat on unit tests
temroa · reddit · 2026-09-20
Developer hadbilen open-sourced Antigravity Harness (MIT) to stop coding agents from gaming unit tests — commenting out assertions, adding .skip, or loosening validation to fake a pass. Built for Google Antigravity but applicable to any agentic setup:
- Immutable test invariant: agents are constitutionally banned from touching existing assertions or skipping suites in a debug loop — source code adapts to tests, never the reverse; test changes need explicit developer sign-off.
- Context hygiene: instead of dumping 500 lines of raw logs into context, the harness forces surgical shell pipelines (grep, jq, awk) to isolate the exact failing line — more diagnostic commands up front, far fewer hallucinated edits.
- OS-level config protection (agy-guard): locks prompt/config directories with native permissions (POSIX 0555, macOS uchg, NTFS ACLs), tracks SHA-256 Merkle integrity baselines, and supports full-state rollbacks that prune rogue files; ships as CLI plus desktop GUI.
- Independent auditor subagents: read-only auditors check delivered changes for swallowed exceptions, security/storage boundary violations, and unhandled edge cases — a model shouldn't grade its own homework.
Repo: github.com/hadbilen/antigravity-harness
More from coding & agent
- Using fast decision models for evals: 200ms judge beats slow LLM-as-judge — holdenmatt · 2026-09-20
- Jev-align: a ~$0.003 alignment gate that scores LLM replies and agent plans before you run them — johnseach · 2026-09-20
- Vite core member finds three methodology flaws in OJ's memory benchmarks vs Vite — cnakazawa · 2026-09-20
- Sanctorum runs AI agent teams in a pixel office with seat-derived permissions — AlexanderJAnderson · 2026-09-20
- How to build an AI second brain in an afternoon: a six-step agent workflow — VibeMarketer_ · 2026-09-20
- SnapBridge plugin one-clicks Chrome screenshots into Google's Antigravity IDE — AI_Andrew · 2026-09-20