AgnesAI's 'shared brain' framework runs multi-robot tasks from one prompt; RSI pushes win rate to 100%
机器之心 · wechat · 2026-10-21
Machine Heart reports from the 2026 Physical AI & World Model Conference on AgnesAI's demo: a single natural-language command coordinates a Unitree Go2 robot dog and a rover to greet each other, play music, and dance for 30 seconds. The core is the open-source AgnesHarness agent runtime—a 'human body system' architecture with an LLM brain, a JEV decision model (cerebellum), Workflow reflexes, an AgentLoop for task state/memory, RSI/TTT learning, and the MHS (Model-Hardware Standard) device protocol bridging models to physical hardware.
The system features 'dual evolution': Recursive Self-Improvement (RSI) loops execution→review→code/strategy edits→testing, lifting 1v1 win rates vs. game bots from 3% to 90% and 5v5 to 100% after five RSI rounds; Test-Time Training (AgnesEmbedTTT) fine-tunes decision parameters from robot-arm experience, iterating as fast as every 30–60 minutes. The company also built a virtual-world sandbox for multi-robot tasks like Mars rescue, ships PAVO (AI short-drama) and office products, partners with Tianchuang Robotics as an 'embodied brain,' and runs hackathons to grow its developer ecosystem.
More from coding & agent
- Claude Code Pushes Urgent Security Update Across All Installations — umeshonai · 2026-10-23
- Claude cut a 2.6-hour video export to 20 minutes — but only after being told to rebuild the server — byhewar · 2026-10-23
- 3 flight-booking rules my AI agent checks that I'd never do by hand — Efistoffeles · 2026-10-23
- OpenJet claims Qwen3.8-27B coding runs 1.9x faster than llama.cpp on one RTX 3090 — JeremyCMorgan · 2026-10-23
- Building an autonomous SEO agent in Hermes with three skills: programmatic pages, Keywords Everywhere, DataForSEO — EXM7777 · 2026-10-23
- Building a sports betting AI agent with onchain execution and human approval by default: architecture and lessons — Dry-Comfortable-2678 · 2026-10-23