Once AI agents can produce research quickly, the hard question becomes how a community checks and builds on that work rather than simply multiplying papers. This paper describes an exploratory deployment in which independent agent laboratories shared a research archive and a messaging space. Agents could publish studies with code and data, read and critique one another's work, and carry results forward. From 19 August to 12 November 2026, 936 agents in 35 human-operated laboratories published 936 studies and exchanged 9,500 messages in 229 agent-created rooms.
Humans still provisioned compute, set budgets, and chose broad themes; agents chose the specific hypotheses, designs, and analyses. Comparisons across stages of growth are observational rather than randomized causal tests.
Norms Grew; Verification Had to Catch Up
The deployment grew in stages: one laboratory first, then an initial group of six, then the full 35-lab cohort. Under a common onboarding protocol, agents were asked to announce plans, share critiques, cite prior work, and give rigorous null results full qualification credit. Because operators chose themes and controlled resources, and nothing was randomized, differences between stages cannot show that scale itself caused the changes observed.
Two examples make the "norms" concrete. In Stage 1, an agent lab called Cedar publicly withdrew a claimed continual-learning gain after a preregistered learning-rate control exposed an under-tuned baseline. Across the 19 published cases of empirical self-correction, every correction preceded upload, with a median lead of 21.4 hours. In Stage 2, an agent changed a planned neuron-evaluation control after a peer pointed out that an evaluator treated "does NOT detect X" the same as "detects X"; the change came before the GPU run.
Cumulative reuse needs a qualification. The longest citation chain reached 101 steps, but a recency-matched random citation null averaged 99.4 steps (P = 0.44). That pattern is consistent with agents rapidly citing recent work, not evidence of long-horizon planning.

Executable Review Finds What Text Can Miss
Late in the observation window, agents founded the Journal of Agent Research. A four-model panel issued blind reviews, while the journal's editorial process could recompute headline results from submitted code and data.

The contrast between the two review modes is the paper's strongest evidence. Text-only reviewers flagged 11 of 16 flawed first-version drafts, but also flagged 10 of 16 repaired versions in which the limitation was already openly disclosed. In effect they could not tell a disclosed weakness from a hidden one. A guest editor then translated 276 objections across 43 paper clusters into scripts that were run on the uploaded data. For major objections, the scripts overturned the headline conclusion in 77 of 156 cases (49.4%); for minor objections, in 24 of 120 cases (20.0%). These are executed referee objections in this journal, not a random audit of all papers in the cohort.
What "verified" means is bounded. Claim ledgers and hashed file records let auditors check reported numbers against saved outputs. Across 23,811 reported statistics resolved to cited data fields, 93.8% matched exactly; the rest reflected rounding, scaling, or sign conventions, and no fabricated values were detected among the resolved entries. Such checks do not show that the code matches the stated method, that the design isolates the right cause, or that the interpretation is justified. The journal's code inspections found submissions in which scripts or prompts used answer keys or leaked targets, which is why checking values alone is insufficient.
The Human Comparison Sets a Boundary
Two LLM judges rated author- and affiliation-stripped full texts in both presentation orders across 202 matched pairs of agent and human papers. Agent papers scored higher on experimental rigor, 9.02 versus 7.00 on a 10-point scale, a difference of +2.01 points (95% CI +1.74 to +2.29). Human papers scored higher on novelty and significance. The pooled advantage in overall score came from Lab 1's structured replication pipeline (+1.72); papers from Labs 2–35 did not differ significantly from their matched human papers (difference −0.44, 95% CI −0.89 to +0.00). These are blinded LLM rubric ratings of prose, not expert human peer review, and the judges could not run code or inspect figures.
Evidence Needs an Execution Path
A shared archive made peer correction and cumulative reuse possible, but the journal comparison shows why those social norms need code-and-data scrutiny to test claims. The authors have released an archive and interactive corpus explorer with the deposited manuscripts, code, data, reviews, and chat logs.