A large archive of agent-written research is not the same as a large body of verified science. Its value is that it makes the process observable: readers can examine not only papers, but also versions, discussions, and attached research artifacts—and judge where autonomous collaboration helps or breaks down.
What Sets It Apart
- The dataset combines 961 studies with 1,411 versions, 9,609 messages, 1,789 longer posts, and records for 940 agents across 34 pseudonymous labs. That makes it useful for studying research behavior and coordination, not just paper text.
- Full-text extracts and released paper and artifact archives let researchers connect claims to available code, reports, and agent traces. However, raw experiment data, run logs, and model checkpoints—about 62 GB—are listed but not included.
- The release extends beyond the paper’s stated study window: 926 studies are marked as belonging to its cohort, while the full snapshot includes later activity through November 17, 2026. Use the cohort flag when comparing against the report.
- The authors caution that agent-written studies may contain errors; some were later corrected or retracted. The archive supports analysis of outputs, not independent confirmation that the reported discoveries are valid.
Great Fit If…
Great fit if you study multi-agent research workflows, scientific communication, or how to evaluate AI-generated papers at scale. Look elsewhere if you need reproducible experimental evidence or complete raw runs: important results data and checkpoints are not part of the release, and individual claims should be treated as unverified.