Microsoft Research has published Agensh, a harness that runs up to 1,024 AI coding agents with no central orchestrator, in a paper first posted to arXiv on 22 September 2026 and revised on 25 October 2026, read on 25 October 2026. The authors report that scaling from 1 to 128 agents “raises the mean final test-pass rate from 19.31% to 28.78%, an approximately 49% relative improvement” on the five hardest ProgramBench tasks.
On one of them, pandoc, the paper says scaling from 1 to 1,024 agents “raises the final test-pass rate from 33.89% to 55.06%”. The code is public on GitHub under an MIT licence. Three days before the revision, on 22 October, Anthropic told companies to fold subagents back into one main loop, as our Anthropic subagents piece reported. The paper adds agents, while Anthropic’s article advises using fewer subagents; the two describe different kinds of work, as set out below.
Business Pill · THE MODEL AND THE HARNESS
A short Business Engineer explainer on the difference between an AI model and the harness around it. It teaches the general idea only and says nothing about the Agensh paper.
The key insight: As we read it, the paper adds a second dial next to the model: with GPT-5.6-sol (high) held fixed, the number of cooperating agents moved the five-task mean from 19.31% to 28.78%, per the paper.
ProgramBench asks agents to rebuild a piece of software’s behaviour from scratch. The paper says it gives them “a 6h budget, with Internet access disabled to prevent retrieval of existing implementations.” The authors picked “the five most difficult tasks among ProgramBench’s 200 instances”: FFmpeg, gromacs, pandoc, PHP-src and ctags.
All runs used the same model, GPT-5.6-sol (high), on top of the Copilot single-agent harness, the paper says. Only the number of agents changed.
Agents
Mean final test-pass rate, five hardest tasks
pandoc only
1
19.31%
33.89%
8
20.68%
—
32
26.52%
—
128
28.78%
50.94%
1,024
not reported
55.06%
Source: Zhan et al., ‘Agensh: Scaling Organizational Intelligence to 1,024 Agents’, Microsoft Research, arXiv 2609.26781. Dashes mark sizes the paper does not report for pandoc.
The paper also says larger teams got there sooner: “on pandoc, 128 agents exceed a 30% test-pass rate at the 30-minute checkpoint, while 32 and 8 agents first exceed that threshold at the 60- and 90-minute checkpoints, respectively.”
Final test-pass rate on pandoc, one of ProgramBench’s five hardest tasks, with 1, 128 and 1,024 agents in Microsoft Research’s Agensh harness, per the paper (arXiv 2609.26781). Same model, GPT-5.6-sol (high), six-hour runs.Mean final test-pass rate on ProgramBench’s five hardest tasks with 1, 8, 32 and 128 agents in Microsoft Research’s Agensh harness, per the paper (arXiv 2609.26781): 19.31%, 20.68%, 26.52% and 28.78%. Same model, GPT-5.6-sol (high), on the Copilot harness, six-hour runs.
How It Works Without an Orchestrator
Most multi-agent harnesses route work through one lead agent, and the paper says their scalability “is often constrained by a central orchestrator’s capacity to allocate tasks and coordinate workers.” In Agensh, every worker runs the same loop, “continuously gathering context, claiming and self-assigning sub-tasks, taking action and sharing findings, verifying results, and merging progress in an asynchronous manner.”
Three shared pieces hold the team together, the paper says: “a shared workspace holds proposed, ongoing, and completed work; a message interface lets workers communicate; and shared context retains reusable findings and work intentions.”
The authors write that roles appeared on their own. All workers “receive the same prompt except for worker IDs”, yet at 1,024 agents “multiple workers take on the same specialized roles”, with several acting as integrators on pandoc.
The Collision With Anthropic’s Advice
Our Anthropic subagents piece reported Anthropic’s 22 October article advising companies to fold routers, planners and subagents back into one main loop. Anthropic said one customer saw “a roughly 2x faster runtime” with a single agent loop, and it listed cases where subagents still help, including when tasks need to run in parallel with fresh context windows.
The two pieces of work measure different jobs. Anthropic describes enterprise deployments, the per-task workflows its Applied AI team saw. Microsoft Research measures long, six-hour attempts to rebuild large codebases, and it treats agent count as a way to add parallel work. The paper does not test Anthropic’s advice, and Anthropic’s article does not discuss Agensh.
Connecting the Coverage
Microsoft chief executive Satya Nadella wrote on 21 October that Microsoft runs a system where the harness, “including context, memory, and action space”, is separate from the model, our Nadella harness piece reported. In the Agensh paper, Microsoft Research also held the model fixed and changed only the number of agents around it.
Yesterday, our company harness piece reported Harrison Chase writing that LangChain’s internal agents “landed on the same core primitives”, five of them. And our Cornell agents piece reported researchers finding that agents took unsafe actions in 1,244 of 1,920 runs after ordinary errors.
What is new here: a controlled test, with the model and the base harness held fixed, showing scores rising as the number of cooperating agents grew to 128, and to 1,024 on one task.
The Structural Read
As we read it, the design choice is the missing boss: the paper says a central orchestrator’s capacity limits current multi-agent harnesses, and Agensh replaces it with a shared workspace, a message interface and shared context.
As we read it, the paper also measures time, not only quality: on pandoc it says 128 agents passed 30% at the 30-minute checkpoint, while 8 agents needed 90 minutes.
As we read it, Anthropic’s 22 October advice and this paper describe different jobs: Anthropic wrote about enterprise deployments its Applied AI team saw, while the paper measures six-hour attempts to rebuild large codebases.
Zhan et al., ‘Agensh: Scaling Organizational Intelligence to 1,024 Agents’, Microsoft Research
“These results demonstrate the number of agents as a new scaling dimension for multi-agent organizations.”
Three Implications
FOR TEAMS RUNNING CODING AGENTS The paper reports higher test-pass rates as agents were added, on its five hardest tasks, with one model; it gives no cost figures.
FOR HARNESS BUILDERS The paper says Agensh coordinates through a shared workspace, a message interface and shared context rather than a lead agent, and the code is on GitHub.
FOR TEAMS WEIGHING ONE LOOP OR MANY Anthropic advised folding subagents for enterprise workflows on 22 October, as we reported; the paper tests long, parallel software rebuilds.
The framework puts it as “The surface, not the model, governs autonomy. The agentic surface — the harness — is becoming where value pools in the entire AI stack.”
As we read it, the paper sits on that surface: it held the model fixed and changed only the number of agents in the harness, and reports the test-pass rates that followed. The framework was published in July 2026 and does not discuss this paper.
What Is Not Established
We read the paper in full; it was first posted on 22 September 2026 and revised on 25 October 2026, and we have not compared the two versions. The results come from one model, GPT-5.6-sol (high), on five ProgramBench tasks, and the 1,024-agent run covers pandoc only. The paper reports no figures on the compute or money cost of large teams. We did not run the code or contact Microsoft Research or Anthropic.
This Week In The AI Supercycle
Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls
This week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.
Microsoft Research reports that a harness with no central orchestrator lifted the mean test-pass rate on ProgramBench’s five hardest tasks from 19.31% with one agent to 28.78% with 128, and pandoc’s from 33.89% to 55.06% with 1,024. Its revision comes three days after Anthropic advised folding subagents into one loop for enterprise workflows.
A note on sourcing. This piece rests on Microsoft Research’s Agensh paper (arXiv 2609.26781), which we read in full, and the paper’s public GitHub repository. Anthropic’s advice and the other earlier facts come from our own reporting, linked in the text. We did not contact Microsoft Research or Anthropic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
