Agent swarm without a central planner lifts pandoc pass rate to 55%
single source· 1 articles · confidence: medium · first seen 2026-09-21 20:00 UTC
What this means for you
Nothing to act on yet. The abstract points to no released code, and no independent group has reproduced the numbers. The transferable observation is that adding agents, not changing models, moved the pass rate — relevant only if you already pay for parallel workers and have hard time budgets.
A preprint reports Agensh, a multi-agent coding harness (the scaffolding that lets a model run code, call tools and check results) built without a central orchestrator. Workers claim sub-tasks from a shared workspace, act, verify and merge asynchronously. On the five hardest ProgramBench tasks, run with GPT-5.6-sol at high effort, going from 1 to 128 agents raised the mean final test-pass rate from 19.31% to 28.78%, a 49% relative gain. On pandoc, 1 to 1,024 agents raised the final test-pass rate from 33.89% to 55.06%. The paper is not peer-reviewed and gives no evaluation date.
Models in this story
Key facts
- ·Agensh is a multi-agent harness with no central orchestrator; concurrent workers self-assign sub-tasks from a shared workspace. source
- ·On the five hardest ProgramBench tasks with GPT-5.6-sol at high effort, scaling from 1 to 128 agents raised the mean final test-pass rate from 19.31% to 28.78%, a roughly 49% relative improvement. source
- ·On pandoc, scaling from 1 to 1,024 agents raised the final test-pass rate from 33.89% to 55.06%. source
- ·The paper is arXiv:2609.26781, posted 21 September 2026. source
What the sources say
- Hugging Face Daily Papers (research) — Single preprint abstract giving scaling curves only; no peer review, no code link, no comparison with other harnesses.
Sources
The original reporting. Follow these — they did the work.
- Hugging Face Daily PapersAgensh: Scaling Organizational Intelligence to 1,024 Agents2026-09-21