rekursiv.ai details how its agent swarm hit SoTA on NanoChat with a 15k-node knowledge graph
hyperparticle · x · 2026-09-15
rekursiv.ai published a full writeup of its auto-autoresearch system: a swarm of AI agents builds on each other's discoveries, preserves failed experiments, and improves how it does research—surpassing the previous NanoChat state of the art in three days.
Setup: improve a language model within a 5-minute training budget on a single B200.
Result: 0.88779 mean bits per byte across 10 seeds, beating the prior best (Recursive 0.91087, Aster 0.90980, Hiloop 0.90160, Hyra 0.90154).
Method:
- Agents treated as "intern savants"—genius knowledge but no sense of what matters—so the swarm figures out the research strategy itself
- A 15,099-node knowledge graph built in 3 days connects issues, artifacts, experiments, beliefs, and code changes
- Built on in-house tools trackinizer, sagent, priml, and configgle
The team previously ran ARC-AGI and Sudoku campaigns and is hiring ([email protected]).
Related event: AI Agent Swarm Breaks NanoChat Benchmark SoTA in 3 Days(4 posts)→
More from coding & agent
- Proposal: hand predictable UI workflows to a local runner to cut computer-use agent costs — Iinganouo · 2026-09-16
- Tutorial: Tracing agent invocations and LLM calls with OpenTelemetry — dl_weekly · 2026-09-16
- Atria Dawn Preview open-sourced: 744B MoE agent model tops BrowseComp at 92.5 — alexcovo_eth · 2026-09-16
- LangChain makes every Managed Deep Agent an MCP server callable by coding assistants — LangChain · 2026-09-16
- LlamaIndex on Building SDKs with Stainless, Whose Team Joined Anthropic — llama_index · 2026-09-16
- HappyRobot CEO on what it takes for an SE to become an FDE at his $1.2B startup — ivory_tang · 2026-09-16