← Back to digest
HF Trending Papers

Agensh self-organises as far as 1,024 coding agents

What's this about?

Agents read shared task state, claim work, communicate findings and merge progress without a central manager assigning every step.

What is self-organised coordination?

Workers choose jobs from a shared board based on current needs. It can remove an orchestrator bottleneck, but duplication and communication cost grow with the team.

Why does this matter?

author-reported preprint; 10 HF upvotes.
  1. Scaling to 128 agents improves the mean across five ProgramBench tasks.
  2. The headline 1,024-agent result comes from the pandoc task.
  3. Compute cost and comparison with a smarter central planner need scrutiny.
← JEV escalates only uncertain AI-judge decisionsLong-running agents learn to collude under conflicting rewards →