Scaling Discovery Does Not Require Forgetting: Verified Shared Archives Break the Multi-Agent Plateau on Anthropic’s Performance Take-Home

Tan et al. report that longer, parallel, and communicating agent searches “all … eventually plateau as the agents accumulate history” and that “the key to scaling discovery further is to forget,” which for them means discarding every agent’s session and the candidate history and restarting from the best candidate. Their experiments never reset sessions while keeping candidates and cannot tell which of the two causes the plateau. We introduce a shared verified archive that bounds each agent’s conversation and keeps every evaluator-verified attempt with its source and cited tools, where any agent can run the attempt or its tools. On Anthropic’s performance take-home, Tan et al.’s forgetting team reaches 1,077 cycles and CORAL 1,103, both with other models. Four Claude Opus 5.5 agents on the archive reach a median of 914 [913, 916] cycles over 3 seeds in 8 hours, against 948 for a CORAL-faithful team and 946 for the best of four isolated agents on matched machines. After 3 hours both baselines’ medians stall between 946 and 956 cycles. Several CORAL-faithful agents concluded that about 951 cycles was near the floor, and 7 of 12 fell into a no-op resume loop. On matched machines the archive’s scores went below that false floor. From 2.75 hours on, the archive’s worst seed stays ahead of both baselines’ best seeds. On open problems an archive team built 60 of its first 98 new bests on another agent’s attempt or tool, and in one run per arm it was ahead of every other arm at 8 hours on AHC026 and on Erdős’s minimum overlap problem and the first autocorrelation inequality.