Microsoft Research: periodically forking an agent into fresh-context copies of its workspace beat single runs by 66% at equal compute

The authors ran autoresearch loops, where a coding agent proposes a solution, scores it and tries again, on 13 long tasks with a deterministic grader. Runs lasted from hours to as long as 107 hours. Independent runs of the same agent on the same task plateaued at very different scores. On one AtCoder-style task (p171), eight runs finished between 68.3 and 100.0. When the authors mapped the submissions by how similar their approaches were, each run stayed in one region, which they call an "idea basin". They report that this held even with periodic compaction of the chat history.

Their fix, fork-and-flush, runs four copies of the agent from the current workspace, each starting with an empty chat. After a short exploration budget, the highest-scoring copy continues and the others are discarded. In the main test the first round starts from empty workspaces, and two more fork-and-flush rounds fire at 33% and 66% of the budget. Under equal compute, the normalized average score was 0.78 for fork-and-flush, 0.47 for a single run and 0.54 for best-of-4, which the authors report as relative gains of 66.0% and 44.4%. Fork-and-flush had the highest mean on 12 of 13 tasks. The exception was LWE-rec, where the single-run baseline's mean was 155.3 against 144.6.

In an ablation over 11 tasks, flushing alone scored 0.55 and forking alone 0.57, against 0.45 for a single run and 0.71 for both together. All conditions used the GitHub Copilot CLI, with GPT-5.6-terra on eleven tasks and GPT-5.5 on the other two. Per-run budgets ran from 6,000 to 100,000 credits, which the paper puts at roughly 100 credits per US dollar. We read the main text and did not read the appendices, and we found no code release.

A second paper, "Scaling Discovery Requires Forgetting" (Shangyin Tan and colleagues, arXiv 2610.20563, October 14), reaches a similar conclusion by a different route. Its abstract says discarding every agent's session and candidate history and restarting from the best candidate so far gave up to 15% over parallel agents and up to 39% over a single agent on four tasks. We have read only that abstract, so treat those figures as the authors' claim.

Why it matters: If you run a coding agent in a score-driven loop for hours, the evidence here says the cost of staying in one long session is a lower ceiling, and that compaction did not fix it in the authors' tests. The tested remedy needs a workspace that survives a context reset and a grader that can pick the winner, which is what a verification loop already provides. The results come from research and optimization tasks, not from ordinary feature work, so the transfer is untested.