Abstract
Searches by teams of language-model agents plateau as agents accumulate history, and Tan et al. propose that agents periodically forget all but the best candidate. Forgetting raises the final best score by up to 15% over agents searching in parallel, but it deletes two things this comparison never separates: the scored attempts and each agent’s conversation, with conclusions that nothing has checked. We introduce the shared verified archive, a multi-agent harness that keeps every scored attempt and lets any agent copy it and run its tools. Conversations stay short: after at most three prompts to continue, an agent starts a new conversation from a summary of the archive. In the first matched 8-hour comparison of this design against parallel and communicating search, four Claude Opus 5.5 agents on the archive reach a median of 914 cycles on Anthropic’s performance take-home (fewer is better). A faithful port of CORAL, whose agents share attempts and notes, ends at 948 and the best of four isolated agents at 946; both medians plateau from the third hour. Several of the port’s agents conclude that about 951 cycles is near the best possible score and stop submitting attempts. We call this belief a false floor, because the same model on the archive beats it. Most of the archive’s improvements after it takes the lead build on a teammate’s attempt, so scaling discovery does not require forgetting scored attempts.
Code and data
Per-seed anytime curves
evidence/analyst_budget_curve.csvTake-home best-so-far cycles per attempt for every seed of the main comparison, the original run, the added seed and the c2d-highcpu-56 rerun70 KBfigures/hero_data.csvFigure 1 data: the nine main-comparison take-home runs and the open-problem arms, every best-so-far step66 KBevidence/curves_matched.csvOpen-problem best-so-far for every arm and problem on a 0.25-hour grid of matched working hours435 KBkx/results/curves-completed.csvTake-home best-so-far curves for the model-change, mixed-family, seeded, open-weight and single-agent runs, with the agent and model of each new best18 KBkx/results/analysis-completed.txtBest-so-far cycles at 15 to 480 minutes for each take-home run in curves-completed.csv, including the mixed-family, seeded, open-weight and model-change runs2.2 KB
The take-home task prompt
kx/harness/task_prompt.mdTake-home task prompt every agent received1.6 KB
C₁ and Erdős constructions with their exact verifiers
anc/ac1-archive-A43.txt.xzon the new-constructions pageC1 construction at 3.97 working hours, 262,144 heights, bound 1.50274043592276697641…1.3 MBanc/erdos-archive-A51.txton the new-constructions pageErdős minimum-overlap construction A51, 3,584 cells, bound 0.38085673888331742055…56 KBanc/verifier/audits_crossfamily/exact.pyon the new-constructions pageIndependent exact checker for C1 and Erdős (GMP carry-free packed convolution, exact fractions)8.2 KBanc/verifier/verify_all.pyon the new-constructions pageRuns every checker on the construction files and compares with the papers' values6.1 KBanc/SHA256SUMSon the new-constructions pagesha256 of every construction and checker file40 KBanc/README.txton the new-constructions pageFile formats, exact values and how to run the checkers6.2 KB
Cite
@misc{cumes2026scaling,
title = {{Scaling Discovery Does Not Require Forgetting: Verified Shared Archives Break the Multi-Agent Plateau on Anthropic's Performance Take-Home}},
author = {Cumes, Lindell S.},
year = {2026},
url = {https://durableintel.com/paper-societies-horizon}
}