Agensh: Scaling Organizational Intelligence to 1,024 Agents

What it claims

Agensh is a multi-agent harness that drops the central orchestrator entirely, letting concurrent workers self-organize. The benchmark setting the summary omits: results come from the five hardest ProgramBench tasks run with GPT-5.6-sol (high). Scaling 1→128 agents lifts mean final test-pass rate 19.31%→28.78% (~49% relative gain); on pandoc, 1→1,024 agents lifts it 33.89%→55.06%. The stated motivation is latency — reducing wall-clock time on complex tasks under hard time budgets by running work concurrently, not just improving accuracy.

How it works

Each worker runs the same cooperation loop rather than receiving assignments from an orchestrator:

flowchart LR
  A[gather context] --> B[claim / self-assign sub-task]
  B --> C[take action + share findings]
  C --> D[verify results]
  D --> E[merge progress]
  E --> A

The loop is backed by three shared-infrastructure components (all named in the summary but worth pinning to their roles): a shared workspace tracking proposed/ongoing/completed work, a message interface for worker-to-worker communication, and shared context retaining reusable findings and work intentions. Coordination is asynchronous — the abstract's point is that this removes the orchestrator as a scaling bottleneck.

Evidence

Two data points, both above. The load-bearing additions beyond the raw numbers: the ProgramBench figures are means over only the five hardest tasks (not a broad suite), and the paper reports that larger organizations reach comparable pass rates earlier — i.e. the gain is partly speed-to-result, supporting the latency framing. The authors also claim that from worker trajectories, distinct forms of self-organized cooperation "gradually emerge and standardize" as agent count grows — an observational finding, not a controlled result.

Limits

This is an abstract-only page; no code, weights, run commands, per-task breakdown, or cost/latency figures are given, and the emergent-cooperation claim is described qualitatively without metrics. The strong scaling numbers rest on a single model (GPT-5.6-sol high) and a narrow task set — the five hardest ProgramBench tasks, with the 1,024-agent result shown only on pandoc. Whether the pattern generalizes beyond these, or what the token/compute cost of running 1,024 concurrent agents is, cannot be judged from the abstract.