Agensh drops the central orchestrator. Workers claim tasks from a shared workspace, and scaling to 1,024 agents lifted pandoc test-pass from 33.9% to 55.1%

arXiv 2609.26781: Agensh·medium signalXBlueskyLinkedInCopy link

arXiv 2609.26781 (22 Sep) replaces the orchestrator bottleneck with three shared pieces: a workspace of proposed, ongoing and completed work, a message interface, and shared context for reusable findings and intentions. Workers gather context, self-assign, act, verify and merge asynchronously. On the five hardest ProgramBench tasks with GPT-5.6-sol, going from 1 to 128 agents raised mean pass rate from 19.31% to 28.78%. For builders, a shared work board that agents claim from is a concrete alternative to a planner that hands out tasks, and it matters most when you have a latency budget.