Open-source AgnesHarness gives agents dual-system brains, cutting task time 32.9%
量子位 · wechat · 2026-11-20
QbitAI profiles AgnesHarness, an open-source system (Apache 2.0, 2.1k GitHub stars) targeting the biggest cost in enterprise AI agent deployment: while development costs collapsed from 90 units to 1, per-customer adaptation still costs 10.
Key designs:
- Sys1 dual-system decision-making: complex reasoning goes to the LLM; frequent, well-defined decisions go to a 9B model, AgnesSys1.0. It decided next steps 14.3x faster than the LLM (0.53s vs 7.6s) at 1/27 the per-decision cost, and cut total task time 32.9% in a robot-dog door-traversal sim.
- TTT × RSI self-evolution: test-time training temporarily tunes a small set of parameters per task (reverted afterward), while RSI post-task reviews distill verified improvements into versioned Playbooks. A device-intake task went from a 30-minute timeout on first run to 8 minutes on the second.
The system is plugin-based (tools, workflows, skills; MCP integration) and adds MHS, a model-hardware standard connecting agents to arms, sensors, and robots, with physical safety left to device controllers. Self-reported results across 9 benchmarks: 6 first places, 1 tied with Codex, 87.07 on DeepPlanning, 40.69% pass on BigCodeBench-Hard V1.
More from coding & agent
- AI turns Mediterranean illustration into interactive 3D world; Three.js wall-crash demo costs $6.98 — Saboo_Shubham_ · 2026-11-20
- GitHub adds reachability analysis to Dependabot alerts to prioritize real vulnerability risk — 0xkarasy · 2026-11-20
- 17 interactive demos in one day with Gemini + Antigravity: 762 tool calls, ~71M tokens, 93% cached — Saboo_Shubham_ · 2026-11-20
- lumimax turns Karpathy's LLM Wiki idea into an MCP knowledge base shared across agents — xiaohu · 2026-11-20
- KOL tests lumimax enterprise knowledge base on 40 NVIDIA filings: conflict detection plus MCP agent access — xiaohu · 2026-11-20
- Dimillian shows new Codex experience running on Duo, driving an autonomous iOS release — Dimillian · 2026-11-20