Microsoft’s Agensh Deletes the Orchestrator: 1,024 Self-Organized Agents Rebuild Pandoc

Multi-agent frameworks have a scaling ceiling that is rarely discussed: the orchestrator. The supervisor that hands out subtasks, tracks progress and reassigns work becomes the bottleneck long before the workers do. More agents on a centralised harness means more coordination traffic through a single point.

Microsoft’s answer is to remove that point entirely.

A Harness Without a Boss

Agensh (arXiv:2609.26781, “Scaling Organizational Intelligence to 1,024 Agents”) is a self-organised multi-agent harness. Workers run a cooperation loop concurrently: gather context, claim and self-assign subtasks, act, share findings, verify each other’s results and merge progress — all asynchronously, with no central coordinator.

Three pieces of shared infrastructure hold it together: a shared workspace for proposed, ongoing and completed work (Gitea), a message interface for worker-to-worker communication (Mattermost), and a shared context layer that retains reusable findings and work intentions (adapted from DeLM). The per-agent agentic loop sits on top of the Copilot CLI harness.

The code, launch scripts and a reproduction guide went public on 24 October under the MIT licence.

What Actually Improved

Evaluated on the five hardest ProgramBench tasks with GPT-5.6-sol at high reasoning effort, a six-hour budget and no internet access:

Mean final test-pass rate: 19.31% with 1 agent → 28.78% with 128 (~49% relative)
Pandoc: 33.89% (1 agent) → 50.94% (128) → 55.06% (1,024)
With GPT-6.1-sol: Pandoc 56.19% (1) → 79.59% (32 agents), ~42% relative

Two details matter more than the headline numbers. Larger organisations reach comparable pass rates sooner — the same quality arrives earlier in the budget, which matters when the constraint is latency. And worker trajectories show that cooperation patterns emerge and then standardise as the organisation grows: roles, integration workflows and peer review stop being hard-coded and start being improvised.

Why It Matters

The interesting claim is that agent count is a scaling dimension in its own right — not just more tokens or a bigger model. And the release is unusually reproducible: a single-node Docker Compose team setup, a standalone single-agent baseline, and native Copilot ACP sessions, plus a published replay of 1,024 agents working together.

The caveats are equally plain: gains are relative, the evaluation covers five hard tasks, and the whole thing was measured with a fixed six-hour, offline budget. That is a research harness, not a production orchestrator — but it is a credible blueprint for one.