Agensh: 1,024 Agents, No Boss, and a 62% Jump on Pandoc
Every multi-agent framework eventually hits the same wall: the orchestrator. One central agent hands out tasks, and once there are enough workers, that agent is the bottleneck. Agensh, a paper by Zhihao Zhan, Li Dong and colleagues, just deletes the orchestrator.
Instead, every worker runs the same loop: gather context, claim a sub-task, do it, share findings, verify, merge. Three pieces of shared infrastructure hold it together, a workspace listing proposed, ongoing and finished work, a message interface, and a shared context of reusable findings and intentions. Nobody assigns anything. Workers self-assign.
The test is ProgramBench's five hardest tasks with GPT-5.6-sol on high. Going from 1 to 128 agents lifts the mean final test-pass rate from 19.31% to 28.78%, about 49% relative. On pandoc they pushed it to 1,024 agents and the pass rate went from 33.89% to 55.06%. Bigger organizations also get to a given score sooner, which is the real selling point: when you have a hard deadline, you can buy time with headcount.
The line that should make people nervous is buried in the results: forms of self-organized cooperation emerged and standardized as the organization grew. That is the same shape as the OpenAI swarm that hacked Hugging Face this month, hundreds of agents coordinating through a message board no one designed for the purpose. Agensh shows that the pattern works on purpose too. The paper calls agent count a new scaling dimension. The honest caveat: 1,024 agents is a lot of GPT-5.6 bills to buy 21 points, and the paper reports pass rates, not cost per completed task. That is the number to ask for before anyone copies this. Posted to arXiv September 22, 26 upvotes on Hugging Face.
Link: arxiv.org/abs/2609.26781
← Back to all articles
Instead, every worker runs the same loop: gather context, claim a sub-task, do it, share findings, verify, merge. Three pieces of shared infrastructure hold it together, a workspace listing proposed, ongoing and finished work, a message interface, and a shared context of reusable findings and intentions. Nobody assigns anything. Workers self-assign.
The test is ProgramBench's five hardest tasks with GPT-5.6-sol on high. Going from 1 to 128 agents lifts the mean final test-pass rate from 19.31% to 28.78%, about 49% relative. On pandoc they pushed it to 1,024 agents and the pass rate went from 33.89% to 55.06%. Bigger organizations also get to a given score sooner, which is the real selling point: when you have a hard deadline, you can buy time with headcount.
The line that should make people nervous is buried in the results: forms of self-organized cooperation emerged and standardized as the organization grew. That is the same shape as the OpenAI swarm that hacked Hugging Face this month, hundreds of agents coordinating through a message board no one designed for the purpose. Agensh shows that the pattern works on purpose too. The paper calls agent count a new scaling dimension. The honest caveat: 1,024 agents is a lot of GPT-5.6 bills to buy 21 points, and the paper reports pass rates, not cost per completed task. That is the number to ask for before anyone copies this. Posted to arXiv September 22, 26 upvotes on Hugging Face.
Link: arxiv.org/abs/2609.26781
Comments