The Pulse
Microsoft Research Tests Agensh With 1,024 AI Agents
A Microsoft Research paper reports that Agensh raised pandoc’s test-pass rate from 33.89% with one agent to 55.06% with 1,024 agents, while the mean result across five ProgramBench tasks rose from 19.31% to 28.78% with 128 agents.
AI.info Team ·
Moving from one AI agent to 1,024 raises the reported test-pass rate on a difficult software-reproduction task from 33.89% to 55.06%, according to a new Microsoft Research paper. The result comes from Agensh, a multi-agent harness designed to let workers organize themselves without a central planner assigning every task.
“A multi-agent system can reduce latency on complex tasks by executing work concurrently.” — Zhihao Zhan, first author, and co-authors at Microsoft Research
The paper, submitted to arXiv on September 22, 2026, presents the system as an answer to a familiar bottleneck in multi-agent software: the orchestrator. In many existing setups, one lead agent decomposes a goal, assigns work to other agents and combines their output. Agensh removes that role and gives workers shared tools for deciding what to do next.
Agensh replaces the central planner with shared infrastructure
Each Agensh worker follows the same five-step loop. It gathers the current project state, claims a sub-task, works with available tools, verifies its result and merges the contribution into the shared workspace. Workers repeat the cycle asynchronously rather than waiting for every other worker to finish.
The system relies on three pieces of infrastructure. A shared workspace stores work in progress and integrated changes. A message interface carries team announcements and direct messages for resolving overlaps or dependencies. Shared context stores reusable findings, failed approaches, active claims and summaries of completed changes.
Microsoft Research implemented the workspace with Gitea and the messaging system with Mattermost. The paper describes Agensh as a layer above an underlying single-agent harness, meaning the model still handles local reasoning and tool use while Agensh manages identity, event routing, shared state, coordination and recovery.
ProgramBench provides the test
The researchers evaluated Agensh on the five hardest tasks in ProgramBench, a software-engineering benchmark that asks agents to reproduce the behavior of reference software from scratch. Internet access is disabled, and each run receives a six-hour time budget. The selected projects are FFmpeg, gromacs, pandoc, PHP-src and ctags.
Those repositories vary widely in size and purpose. The reference codebases cover multimedia processing, molecular simulation, document conversion, language interpretation and code indexing. The largest of the five, PHP-src, contains more than 26,000 files and about 2.8 million lines of code in the version used for the evaluation.
Using GPT-5.6-sol with the high setting, the mean final test-pass rate across the five tasks rises from 19.31% with one agent to 28.78% with 128 agents. The change amounts to a 9.47 percentage-point gain, or about 49% relative improvement.
A larger organization reaches the higher result
The most aggressive experiment scales pandoc from one to 1,024 agents. The single-agent result is 33.89%, compared with 55.06% for 1,024 agents. Under the same six-hour budget, the largest organization improves on the single-agent result by 21.17 percentage points.
Agensh’s evidence comes from one benchmark family, one model configuration and a tightly defined six-hour software task. The experiments show that adding workers can improve final performance, but they do not establish that the same scaling pattern will hold for general-purpose work or for applications where coordination, inference and infrastructure costs matter more than elapsed time.
The paper’s more specific contribution is architectural. It treats agent count as an experimental variable and supplies shared files, messages and persistent findings instead of a single manager. In the reported pandoc run, that design carries the organization from a 33.89% test-pass rate with one worker to 55.06% with 1,024 workers under the same six-hour budget.