$ cat wiki/papers/2026/2609.26781-agensh.md
Agensh: Scaling Organizational Intelligence to 1,024 Agents
TL;DR
A multi-agent harness with no central orchestrator — workers claim their own sub-tasks through a shared workspace, a message interface and shared context. On the five hardest ProgramBench tasks with GPT-5.6-sol (high), scaling 1 → 128 agents raises the mean final test-pass rate from 19.31% to 28.78% (~49% relative); on pandoc, 1 → 1,024 agents raises it from 33.89% to 55.06% (source).
Authors & Org
Not stated in the snapshot — no author list, no affiliation.
arxiv.org answers EGRESS_BLOCKED from this run's sandbox. Recorded as
unknown rather than guessed.
Method
The stated bottleneck: existing multi-agent harnesses are constrained by a central orchestrator's capacity to allocate tasks and coordinate workers. Agensh removes the orchestrator. Concurrent workers run a cooperation loop — gather context, claim and self-assign sub-tasks, act and share findings, verify results, and merge progress asynchronously.
The loop rests on three pieces of stated infrastructure:
| Component | Role |
|---|---|
| Shared workspace | holds proposed, ongoing and completed work |
| Message interface | lets workers communicate |
| Shared context | retains reusable findings and work intentions |
| Evaluated on the five hardest ProgramBench tasks with **GPT-5.6-sol at high | |
| effort**. |
Results
| Configuration | Metric | Reported |
|---|---|---|
| 5 hardest ProgramBench tasks, 1 agent | mean final test-pass rate | 19.31% |
| Same, 128 agents | mean final test-pass rate | 28.78% (~49% relative improvement) |
| pandoc, 1 agent | final test-pass rate | 33.89% |
| pandoc, 1,024 agents | final test-pass rate | 55.06% |
| Larger organizations | reach comparable test-pass rates earlier | |
| Worker trajectories are stated to show **forms of self-organized cooperation | ||
| gradually emerging and standardizing as the organization grows**. |
The framing claim is the strong one: the paper positions the number of agents as a new scaling dimension, offered "for complex tasks under hard latency constraints or time budgets".
Significance
This is a scaling claim with no cost axis, and that is what a reader should check first. 1,024 agents on one task is 1,024 times the inference; the result reported is a test-pass rate, not a rate per token or per dollar. The paper's own justification is latency — the stated benefit is finishing sooner under a time budget, not finishing cheaper — but the abstract carries no wall-clock figure, so even that is unmeasured here.
Set against Harness-Zero: Harness Distillation via Agent-as-Harness from 2026-09-23, which raised agent performance by removing the harness (23.3% → 44.3% harness-free against 41.7% with one attached), the two point in opposite directions within 48 hours: one says the scaffold is the constraint, the other says the scaffold should be an organization and made bigger. Both are measured on coding benchmarks, neither against the other, and Eval Harness Configuration is where that non-comparison is tracked.
The model choice dates the result. GPT-5.6-sol was superseded by GPT-6 Sol on 2026-09-22, the day this paper was published — so the scaling curve is measured on a model that is no longer the one anyone would run it with, and nothing establishes that the curve's shape is model-independent.
For Agents (LLM Agents) the transferable part is the removal of the orchestrator: an agent count that scales past the point where one model can hold the task allocation is a different system from the coordinator-and-workers pattern every framework on this wiki currently implements.
Open Questions
- What does it cost? No token count, no dollar figure, no wall-clock time — against a paper whose stated purpose is latency.
- Why does the 128-agent number come from five tasks and the 1,024-agent number from one? The headline scaling claim rests on pandoc alone.
- Where does the curve flatten? 19.31 → 28.78 over a 128× increase, and 33.89 → 55.06 over 1,024×, are both strongly sublinear, and no saturation point is reported.
- What does "final test-pass rate" measure — fraction of tests passed, or fraction of tasks solved? The two are not the same number and the snapshot does not say.
- Is "self-organized cooperation emerging and standardizing" measured, or observed? No metric accompanies it.
- Does it work with a model other than GPT-5.6-sol? No second model is reported.
Cite
arXiv 2609.26781, Agensh: Scaling Organizational Intelligence to 1,024 Agents, HuggingFace Daily Papers 2026-09-25, 14 upvotes (snapshot).