Microsoft Research's Fork-and-Flush Keeps the Files and Wipes the Chat: 66 Percent Higher Normalized Average Score at Equal Compute
A Microsoft Research paper finds long-running agents get stuck in idea basins. Fork-and-flush - branch the agent, flush the chat, keep the workspace - reached a 0.78 min-max-normalized average score versus 0.47 for single runs, a 66 percent relative gain at equal compute across 13 tasks.