rekursiv's Agent Swarm Breaks NanoChat SoTA in 3 Days, Open-Sources Research Tracking Tools

rekursiv.ai's team claims its swarm of AI agents broke the previous SoTA on Karpathy's NanoChat benchmark in 3 days, followed by a detailed blog post and open-sourced research tracking tools, demonstrating an engineering path for automated AI research.

Confirmed

• The team built an auto-autoresearch harness on top of a graph database, letting agents hand off experiments to each other, claiming to surpass the prior SoTA within 3 days (@hyperparticle, @zsakib).
• The system retains failed experiments and improves its own research process, forming a self-reflective loop; it built a knowledge graph of roughly 15,000 nodes in 3 days (@hyperparticle's blog).
• A real engineering issue emerged: when an AI scientist finished an experiment, the agent awaiting results might no longer exist; the system added an AI reviewer as a supervisor backstop, bridged by the knowledge graph (@hyperparticle follow-up).
• Guided by "what should we test next? what would change our minds?", the team open-sourced three research tracking tools, including trackinizer, for tracking efforts, beliefs and findings of agents and humans, and links between ideas, experiments and evidence (@hyperparticle).

Why it matters

• The work moves automated AI research from concept to a running system, showing agent swarms can execute real benchmark campaigns.
• The knowledge-graph-plus-supervisor design addresses lifecycle and result-attribution problems in long-horizon agent collaboration.
• The open-source toolchain provides ready infrastructure for idea-experiment-evidence tracking, potentially lowering the barrier to reproducing automated research.

2026-09-15 ~ 2026-09-16 · 5 related posts

• [source] AI agent swarm beats NanoChat benchmark SoTA in 3 days with 15k-node knowledge graph — hyperparticle · 2026-09-15
• When the agent awaiting results disappears: supervisor + durable records fix late results — hyperparticle · 2026-09-15
• [source] rekursiv.ai open-sources trackinizer, an epistemological database for agent research — hyperparticle · 2026-09-15
• [source] rekursiv.ai details how its agent swarm hit SoTA on NanoChat with a 15k-node knowledge graph — hyperparticle · 2026-09-15

1 near-duplicate retellings: zsakib_