TL;DR
Microsoft’s open-source Agensh coordinates agents without a central orchestrator. In the authors’ experiments, the mean test-pass rate across five difficult programming tasks rose from 19.31% with one agent to 28.78% with 128. A 1,024-agent result concerns one task, Pandoc, and does not establish lower cost per successful task.
Beyond a manager and its workers
A central orchestrator can decompose a job and hand pieces to individual agents. That design is intuitive, but a single manager can become a bottleneck as a team grows. Agensh explores a different approach: agents discover and claim subtasks, share findings, verify work and merge progress asynchronously. The project is open source under the MIT license. (Source : Microsoft — Agensh repository; Microsoft Research — Agensh paper)
The important design choice is not an absence of coordination. Agensh replaces a central manager with shared infrastructure: a workspace for work in progress, a messaging interface and shared context. In the reported implementation, Gitea provides the workspace and Mattermost supplies messaging. This shifts coordination into protocols, artifacts and peer interactions rather than eliminating it. (Source : Microsoft — How Agensh works)
The five-task result
The authors evaluate the system on five difficult ProgramBench tasks: FFmpeg, GROMACS, Pandoc, PHP-src and ctags. Their setup uses GPT-5.6-sol at high reasoning effort, a six-hour budget and no internet access. Agents reconstruct reference programs using observations and bundled documentation. (Source : Microsoft — Experiment overview; Agensh — Research paper)
Across those five tasks, the reported mean final test-pass rate rises from 19.31% with one agent to 28.78% with 128 agents. That is a 9.47 percentage-point increase, or approximately 49% relative to the starting rate. Relative gains and percentage-point gains answer different questions; reporting both makes the scale of the result easier to interpret. (Source : Agensh — Abstract)
This is evidence of better results under the stated time limit and experimental setup. It is not evidence that the larger team used fewer tokens or cost less. A fixed wall-clock budget allows many agents to run concurrently, so the amount of compute can grow substantially as the team grows.
What happened at 1,024 agents?
The widely cited 1,024-agent number comes from Pandoc, not the average of all five tasks. On that task, the reported pass rate is 33.89% with one agent, 50.94% with 128, and 55.06% with 1,024, all under the six-hour budget. (Source : Microsoft — Agensh results)
The additional improvement from 128 to 1,024 is 4.12 percentage points on this task. That is a useful warning against extrapolating a straight line from team size to quality. It is not a universal scaling law: the shape can depend on the task, model, time budget and coordination design. We also do not have a corresponding cost comparison in those reported headline figures.
The repository separately reports a Pandoc experiment using GPT-6.1-sol: 56.19% with one agent and 79.59% with 32, at high reasoning effort and within six hours. Because the model differs, this is not a clean continuation of the GPT-5.6-sol scaling series. (Source : Microsoft — Agensh repository)
The reproduction question
Open code improves the opportunity to inspect and reproduce the work, but it is not itself an independent replication. Microsoft’s release includes launch scripts and a reproduction guide. The guide also describes differences between the released setup and the historical experiment, including the cleanroom component. Anyone reporting a reproduction should document the actual environment rather than assume it is identical. (Source : Microsoft — Reproduction guide)
A useful follow-up would run repeated trials at several team sizes while publishing total tokens, compute use, elapsed time, test outcomes and variance. Two comparisons would answer different questions: one with a fixed elapsed-time budget, and another with a fixed monetary or compute budget. The first asks how quickly additional parallel workers can improve results. The second asks whether distributing work is more efficient than spending the same resources differently.
Where decentralized teams can still fail
Shared workspaces create new operational questions. Two workers can claim overlapping work, merge incompatible changes or amplify a faulty assumption from shared context. Verification and integration processes become as important as the mechanism for launching agents. Observability must capture who changed what, which tests ran and how conflicting results were resolved.
For a production deployment, a team would also need to test secrets handling, permissions, resource limits and recovery from interrupted workers. The research result motivates those experiments; it does not provide a blanket safety guarantee.
FAQ
Did all five tasks run with 1,024 agents?
The headline 1,024-agent score is reported for Pandoc. The five-task mean comparison uses one and 128 agents.
Does decentralized mean no coordination?
No. Agents coordinate through shared workspaces, messages and context rather than a single central orchestrator.
Is a larger team cheaper?
The headline results establish test-pass rates under a time budget, not a lower cost per successful task.
Further Reading
— The Agent Report