Reverse-Harness lets agents pull tools at runtime under risk governance, benchmarked 6% to 61%
Rameswaranmohan1993 · reddit · 2026-09-09
A solo open-source (MIT) project, Reverse-Harness flips typical agent design: instead of pushing a fixed toolkit, the running agent issues REQUESTHARNESS when it lacks a capability, and a Governor arbitrates by risk. Reusing existing tools auto-grants; generated code always escalates to human review with visible spec and source; dependencies get their own gate; first executions are gated as dangerous-until-proven; every grant is leased, logged, and revocable. Two non-obvious ideas: self-requested capabilities are gated more strictly than pre-provisioned ones, and the LLM judge can only deny or escalate, never grant beyond policy. Across 179 trials, 5 models from 5 vendors, tasks impossible without acquiring a missing capability: 6% with a frozen toolkit vs 61% with governed pull (McNemar p = 2.8e-10). Honest caveats: small models write worse tools, install is heavy, and the eval is the author's own.
More from coding & agent
- Claude helps unlock HDR10 capture from Elgato 4K X on macOS, unofficially — talkaboutdesign · 2026-09-09
- vLLM agentic serving details: packed KV saves 10% memory, Kimi K3 decode throughput up 2.7x — vllm_project · 2026-09-09
- vLLM: agent sessions median 43 turns, 142K-token inputs vs 444-token outputs — vllm_project · 2026-09-09
- vLLM's agentic serving deep-dive: DeepSeek V4 Pro hits 83K tok/GPU-s at 106x lower cost than Opus 5 — vllm_project · 2026-09-09
- vLLM details full-stack optimizations for real-world agentic serving on AgentX benchmark — vllm_project · 2026-09-09
- qacman turns QR codes into playable Pac-Man mazes with custom skins — ezshine · 2026-09-09