On October 24, 2026, Microsoft Research open-sourced Agensh, a harness for large teams of coding agents with no project lead. The published scores are more useful than the headcount. On pandoc, offline and under a six-hour cap, 1, 024 agents using GPT-5.6-sol reached a 55.06% final test-pass rate. The mean across the five hardest ProgramBench tasks, published through 128 agents, rose from 19.31% with one agent to 28.78%. The scores moved inside a fixed budget. The job is not solved.
No project lead
Agensh is a multi-agent harness. It is not a new model. A harness is the wrapper that gives a coding agent tools, files, and a loop. This one sits on top of GitHub’s Copilot CLI rather than replacing the agent runtime.
The preprint, “Agensh: Scaling Organizational Intelligence to 1, 024 Agents, ” went up on September 22, 2026. The authors are Zhihao Zhan, Ting Song, Li Dong, Shaohan Huang, Jianxun Lian, Yan Xia, and Furu Wei. Code, launch scripts, and a reproduction guide followed on October 24 under an MIT license, in the Microsoft Research repository. The project page is at aka.ms/Agensh.
Most multi-agent setups still have a boss. A lead agent plans the work, hands out pieces, and merges the result. That boss becomes the ceiling. It cannot read every file, settle every conflict, or hold a growing team in one context window. Agensh’s bet is to delete that role. Workers discover and claim their own sub-tasks, share findings, verify results, and merge into a common repository, concurrently and asynchronously. Nobody waits for an assignment.
Three services do the connecting. Gitea, a self-hosted Git host, holds the workspace, branches, and merges. Mattermost, a team chat tool, carries the channel and the direct messages. A shared context board, adapted from the DeLM project, keeps short findings so peers do not re-derive a fact someone already confirmed, or burn budget on a hypothesis someone already killed.
Self-organized peer coordination, integration workflows and specialized roles gradually emerge as the team grows.
That line is the repository’s, and it is a qualitative claim. Read it as an observation about how the team behaved, not as a staffing chart you can copy into an org design.
Three boundaries
If you are deciding whether a giant team of AI agents belongs in an automation stack, three limits should sit next to the diagram.
- Gains were uneven, and the average stayed modest. More agents raised the published mean, and that mean is still a long way from done. On pandoc, the only task with a published 1, 024-agent score, the step from 128 agents to 1, 024 is a few points inside the same six hours.
- The clock was part of the result. Every cited score is six hours, offline, at high reasoning effort. A higher score at a larger team means better use of that window. It does not prove the work would keep improving if you left it running all week.
- The bill is missing. Token volume and dollar cost are not disclosed. A swarm that looks clever and costs more than the work it replaces is just a very busy meeting.
The published curve
The test is ProgramBench. Agents rebuild a reference program from observed behavior and bundled documentation, with no internet and a hard stop at six hours. You cannot “open the repo and patch a bug with search a click away.” This is offline reconstruction under a clock. Microsoft labels the metric a final test-pass rate. The formula, the number of trials, and any error bars are not published, so a point is a point on their scale. You cannot recompute it from the README.
On the five-task mean with GPT-5.6-sol, one agent scored 19.31% and 128 agents scored 28.78%. The difference is 9.47 points. The README calls that approximately 49% relative, about a 50% relative increase rather than a 49-point jump. The published mean stops at 128. The 1, 024-agent figure in the repository is for pandoc, so do not read the other tasks as thousand-agent results.
Pandoc, same model and same six hours, runs from 33.89% with one agent to 50.94% at 128 and 55.06% at 1, 024. The last step is 4.12 points. Eight times the team, a few points of pass rate, still inside the cap. The README also says larger teams reach comparable pass rates sooner. It does not print minute-level checkpoints.
The repository overview does not name the five tasks or print per-task scores. Marktechpost, writing the day of the release, names them as FFmpeg, gromacs, pandoc, PHP-src, and ctags, says scores rose on all five, and describes the 1, 024-agent run as 16 nodes × 64 agents with a 64 GiB memory ceiling on each worker. A rising mean does not by itself prove every task improved. Use the README numbers as the result. Treat the rest as secondary operational color the repository overview does not print.
Two pandoc experiments, not one
The repository overview also reports a newer model, GPT-6.1-sol, on pandoc only, again at high effort and six hours: 56.19% with one agent and 79.59% with 32. The README calls the relative improvement 41.64% (relative, not 41.64 points). Those figures shipped with the October 24 code release. The preprint is dated September 22, so do not assume they are the paper’s scaling table.
Set beside the 55.06% from 1, 024 agents on GPT-5.6-sol, the 79.59% looks like a smaller team won. It did not, not in any controlled sense. Model and headcount changed together. The useful reading is narrower. On this one task, the newer model at 32 agents outscored the older model at 1, 024. More AI agents will not fix the result if the model underneath is the binding constraint.
What the public write-ups leave out is just as important for anyone pricing AI for business. No pass-rate formula, no trial count, no variance. No individual scores for the tasks other than pandoc, only the five-task mean. No merge-conflict rate, no measure of duplicated work, no coordination overhead at 1, 024 agents. The names GPT-5.6-sol and GPT-6.1-sol are used as tested. The provider, the availability, and what “sol” means are not explained.
Shared state is the idea worth keeping
The paper title talks about organizational intelligence. The mechanism is humbler. Agensh does not invent a genius manager. It gives workers a place to put code, a place to talk, and a place to leave facts and dead ends, then lets them claim work themselves.
Shared state is the pattern that transfers to AI automation. Headcount is not. A lead agent that plans, assigns, and merges everything will hit an orchestrator ceiling, whether the job is reconstructing a large codebase or splitting a workflow across research, drafting, and follow-up. A flat team still needs shared state. Without a claim on the board and a record of what already failed, you do not get a swarm. You get expensive overlap.
Coverage of the paper points to lead-agent designs (Claude Code agent teams, Codex sub-agents, Copilot’s fleet mode, Kimi Agent Swarm) as the contrast. Marktechpost turns that into a feature table of coordination style, disclosed scale, shared state, and messaging. It is not a ProgramBench bake-off. None of those systems is a measured loser here. Thin adapters, small connectors meant to plug in other harnesses such as Claude Code, are mentioned without published results.
If you want to try the released system rather than the headline run, start small. The repository ships a single-node Docker Compose setup and a single-agent baseline, on Linux, with a model API token. Marktechpost puts the default team at eight workers, six hours, and GPT-5.6-sol at high effort, and cites ProgramBench v1.2.5 as the validated benchmark version. Cloning the repo and hitting run does not reproduce the thousand-agent experiment. Measure cost yourself. Do not staff a thousand agents because the published curve rose for a while.
Questions worth asking before you staff a swarm
-
Does adding agents reliably raise the score?
The published means rose. Reliability was not reported. No trial counts, variance, or error bars. We cannot tell whether that lift would show up again on a rerun, and the average remains well short of a solved task.
-
Is 1, 024 the setup you get from the open-source release?
No. The release is a single-node Compose setup plus a single-agent baseline. The 16-node, 1, 024-agent topology is the experimental run described in secondary coverage, not the documented default.
-
Which pandoc number is comparable to which?
Compare GPT-5.6-sol with GPT-5.6-sol, and keep the GPT-6.1-sol series separate. The 55.06% figure is 1, 024 agents on the older model. The 79.59% figure is 32 agents on the newer one, from the repository overview. They are not the same experiment.
-
Were the specialist roles assigned in advance?
The repository says peer coordination, integration workflows, and specialized roles emerge as the team grows. It does not present them as jobs assigned up front.
-
What did 1, 024 agents cost?
It is not disclosed. Without token volume or dollar cost, the business case cannot be closed from the scores alone.
-
Should you rip out the lead agent this quarter?
Not on this evidence. Agensh is a serious test of whether a shared workspace, messaging, and a findings board can replace a central orchestrator on hard coding work. It is not a bake-off, and the absolute scores stay modest.
The fantasy of a company with no managers and a thousand diligent specialists is old. Agensh is one of the first public attempts to run that fantasy as a coding system and publish the curve. The curve bends. It does not break upward forever. For anyone buying AI agents, that bend is the finding: coordination can emerge without a boss, a better model can beat a bigger team when you change both at once, and the bill for either choice is still unpublished.