Microsoft open-sources Agensh, a 1,024-agent harness

The reproduction guide covers a single-node team, run with Docker Compose, and a standalone single-agent baseline[3]. Both use native Copilot ACP sessions and ProgramBench's v6 cleanroom. The defaults it prints are 8 workers, 6 hours per task, gpt-5.6-sol, high reasoning effort, the Responses API and a 272,000-token prompt window[3].

Release validation used ProgramBench v1.2.5 at commit 27f02157c785f8da3647aa6dbbe6b9137f99f10e[3]. The guide pins Copilot 1.0.70 and Mattermost MCP 0.5.1. The host it specifies is Linux with Docker Engine, Compose v2, systemd and cgroup v2. The README and the guide both say the code is under the MIT License[1][3].

The guide's five-task example uses jgm__pandoc.5caad90, gromacs__gromacs.665ea4c, universal-ctags__ctags.243595e, php__php-src.c891263 and ffmpeg__ffmpeg.360a402[3]. Those short hashes are the commits in Table 1 of the paper[2].

The paper tests Agensh on ProgramBench, where agents rebuild the behavior of reference software from scratch, with internet access disabled, inside a six-hour budget[2]. It selects FFmpeg, gromacs, pandoc, PHP-src and ctags, and it calls them the five most difficult of ProgramBench's 200 instances, ranked by the mean test-pass rate of state-of-the-art models. Table 1 lists PHP-src, commit c891263, at 26,266 files, 2,812,757 lines and 119.76 MB, and pandoc, commit 5caad90, at 2,767 files, 104,336 lines and 4.86 MB[2].

With GPT-5.6-sol at high reasoning effort, Copilot as the underlying single-agent harness, and the same six-hour budget, the mean final test-pass rate across the five tasks moves from 19.31% with 1 agent to 20.68% with 8, 26.52% with 32, and 28.78% with 128[2]. The paper calls the 1-to-128 change a gain of 9.47 percentage points, or an approximately 49% relative improvement. The README prints the same 19.31% and 28.78% endpoints and the same approximate relative improvement[1].

On pandoc, the paper says the final test-pass rate rises from 33.89% with 1 agent to 50.94% with 128 agents and 55.06% with 1,024 agents[2]. It says the 1,024-agent result is 4.12 percentage points above the 128-agent result and 21.17 percentage points above the single-agent result. The project page prints the same 33.89% to 55.06% pair and labels it a "+62% pass rate relative gain"[6].

On timing, the paper says that during the first two hours of the six-hour runs, 128 agents exceed a 30% test-pass rate on pandoc at the 30-minute checkpoint, while 32 agents and 8 agents first exceed that threshold at the 60-minute and 90-minute checkpoints. The single-agent run remains below that threshold throughout the first two hours[2].

The appendix says every reported configuration uses gpt-5.6-sol with high reasoning effort, with a maximal input token limit of 272,000 and a maximal output token limit of 128,000[2]. Activation is staggered: one agent every 30 seconds during the first hour, then one every 3 seconds. In the 1,024-worker experiments, workers are distributed across 16 nodes, with 64 agents per node[2].

The README prints a second pandoc series. With GPT-6.1-sol at high reasoning effort, it says the six-hour test-pass rate rises from 56.19% with 1 agent to 79.59% with 32 agents, a relative improvement of 41.64%[1]. The figure in demo/pandoc_results.png is the chart for that comparison. The README's alt text names GPT-5.6-sol and GPT-6.1-sol at high reasoning effort[1]. The paper's pandoc paragraph reports GPT-5.6-sol, and it follows that model out to 1,024 agents[2].

The guide says the historical main experiment seeded the older task_cleanroom reference, and that its binary differs from v6 even when version strings match[3]. It says these infrastructure updates mean a new release run should record its own revision rather than claim exact historical reproduction[3].

The release ramp targets 30 seconds per worker for the first 120 workers, then 3 seconds, with a serial 20-second settle wait[3]. The guide says actual startup can exceed those targets, so later workers get less time, that the ramp can overrun the budget, and that a minimum 60-second wait follows. Workers have a 64 GiB memory ceiling and routers have 8 GiB. Those figures are ceilings, not reservations. Worker networks restrict external access to model egress[3].

The paper's cooperation loop has five steps. A worker gathers context, claims a sub-task, takes action, verifies the result against the sub-task's acceptance criteria, and merges the contribution into the shared workspace, then starts again[2]. Workers run that loop concurrently and asynchronously. The shared workspace is a Git platform, implemented with Gitea. The message interface is Mattermost: a shared task channel for announcements, and direct messages for urgent one-to-one notes. Shared context, adapted from DeLM, stores short typed entries the paper names OBSERVED, FACT, FAIL, CLAIM and PATCH_SUMMARY[2].

The loop lives in the worker prompt. The paper says the prompt is the same for every worker except the worker ID, and that adapters can place it on an underlying harness such as Claude Code or Copilot without changing the loop, the shared services, or that harness's own agentic loop[2]. The README says the experiments sit on the per-agent loop of the Copilot harness[1].

The paper treats the recorded trajectories as evidence that cooperation changes with headcount, under one shared prompt. At 8 agents it describes workers agreeing an interface and resolving overlapping claims. At 32 it describes peers withdrawing an approval after a counterexample and then merging a fix. At 128 it describes reusable review relationships and a pandoc integration protocol that other workers later pick up. At 1,024 it describes several workers acting as integrators, with a worker contacting more than one and handing code to the first valid responder[2]. Those are the paper's readings of the logs.

v1.0 is a single-node MIT release. Its guide defaults to 8 workers, and the commit that the tag points at describes a single-node harness[3][5]. The 55.06% pandoc figure is the paper's result for 1,024 agents of GPT-5.6-sol, spread across 16 nodes, with no internet and a six-hour budget[2]. The 79.59% figure is the README's result for 32 agents of GPT-6.1-sol on the same task and budget[1]. The reproduction guide tells a new run to keep its own revision, because the cleanroom image has changed since the historical experiment[3].