Microsoft Agensh scaled a software-engineering experiment to 1,024 self-organizing AI agents without a central orchestrator. Microsoft Research reported that the largest group produced a 55.06% test-pass rate while rebuilding Pandoc from observed behavior, compared with 33.89% for one agent under the same six-hour limit.

 

The September 22 paper presents agent count as a new way to improve performance and reduce latency on complex work. The result is notable, but narrow: Agensh remains a research system, the 1,024-agent run covered one task, and the preprint has not undergone independent peer review.

 

The experiment establishes four central findings:

  • Workers chose and claimed their own subtasks.
  • A shared workspace, messaging and memory replaced one central manager.
  • Scaling from one to 128 agents improved the five-task average by about 49%.
  • The 1,024-agent Pandoc run reached a 55.06% test-pass rate.

 

Microsoft Agensh Removes the Central Orchestrator

Many multi-agent systems use a lead model to break down a job, assign pieces to workers and combine their results. That design can become a bottleneck as the group grows because every plan, dependency and contribution must pass through one coordinator.

 

Agensh replaces that hierarchy with a shared operating protocol. Every worker repeatedly gathers context, claims a useful subtask, takes action, verifies its result and merges the contribution. The agents run concurrently and asynchronously, so the group does not wait for a synchronized round to finish.

 

Three infrastructure components hold the organization together. A Git-based workspace stores work and exposes merge conflicts, a Mattermost message layer supports announcements and direct coordination, and an append-only context service records verified facts, failed approaches, active claims and completed patches.

 

The distinction is architectural rather than a claim that coordination disappears. Workers still negotiate overlapping responsibilities, review contributions and resolve conflicts. Agensh distributes those decisions among peers instead of asking one supervising agent to maintain a complete picture of the project.

 

ProgramBench Tests Rebuilding Five Large Software Projects

The researchers evaluated Agensh on the five hardest instances in ProgramBench: FFmpeg, GROMACS, Pandoc, PHP and Universal Ctags. The benchmark asks agents to recreate the behavior of a compiled reference program without internet access, using experiments against the executable rather than reading the original source.

 

Each organization received six hours. Every worker used GPT-5.6-sol at high reasoning effort through the same Copilot-based single-agent harness. Holding the model, tools and time limit constant allowed the team to study how the number of collaborating workers changed the outcome.

 

Across all five tasks, the mean final test-pass rate rose from 19.31% with one agent to 20.68% with eight, 26.52% with 32 and 28.78% with 128. The one-to-128 increase was 9.47 percentage points, or approximately 49% on a relative basis.

 

Larger teams also reached some performance levels sooner. On Pandoc, 128 agents crossed a 30% pass rate after 30 minutes, while the 32-agent and eight-agent configurations first crossed that threshold after 60 and 90 minutes. The single-agent run stayed below it for the first two hours.

 

How 1,024 AI Agents Organized Their Work

The largest experiment focused on Pandoc and distributed 1,024 workers across 16 nodes. Agents were activated gradually to reduce immediate contention. After six hours, the group reached 55.06%, versus 50.94% for 128 agents and 33.89% for the single-agent baseline.

 

The improvement from 128 to 1,024 workers was 4.12 percentage points, much smaller than the 17.05-point gain between one and 128. That pattern suggests continued benefit alongside diminishing returns, although the paper does not provide enough tasks at the largest scale to define a general scaling law.

 

Recorded trajectories showed coordination changing with group size. Eight-agent teams negotiated interfaces and divided components. At 32 agents, peers reviewed and reworked shared contributions. At 128, repeated reviewer relationships and standard integration procedures appeared. With 1,024 workers, multiple agents adopted specialist integration roles and could replace one another after failures.

 

Those behaviors were not assigned as fixed job titles. All workers received the same cooperation instructions apart from their identifiers. The specialization emerged from their messages, claims, review history and access to shared context, which is the study's main evidence for self-organization rather than centrally scheduled parallelism.

 

Agensh Results Need Cost and Replication Tests

The experiment demonstrates that a large peer organization can make measurable progress on a difficult coding benchmark. It does not show that deploying hundreds of agents is economical, reliable across domains or preferable to improving the underlying model. The paper does not present a full cost comparison for the configurations.

 

Its evidence is also concentrated. The one-to-128 results span five demanding programs, but the 1,024-agent result is limited to Pandoc. Program reconstruction under a six-hour deadline rewards parallel exploration, so other work involving sequential decisions, scarce tools or tightly coupled changes may scale differently.

 

Operational risks grow with the team. More workers can duplicate effort, flood shared memory, create conflicting patches and consume model capacity before useful work is integrated. Agensh addresses these problems with claims, concise shared records, messaging and Git history, but the remaining overhead is part of the result, not a solved issue.

 

Microsoft Research released the paper and linked a public code repository, making replication possible. The next useful evidence would include repeated trials, additional benchmarks, token and infrastructure costs, failure rates and comparisons with strong centrally orchestrated systems under matched budgets.

 

If those tests hold, Agensh would shift attention from building a smarter manager toward designing environments where many capable workers can coordinate directly. The paper's clearest contribution is not the headline number alone, but a concrete protocol for turning parallel agent activity into verified, mergeable progress.

 

Related Research