Microsoft researchers found that AI agents often get stuck refining the same ideas because their chat history steers them back into the same solution space. Their new method, called Fork-and-Flush, creates several parallel branches, wipes each branch's chat context, and keeps only the best results. On 13 benchmark tasks, the technique lifted the normalized average score from 0.47 to 0.78 while also reducing result variance.
Why it matters
Developers can obtain higher‑quality, more consistent solutions from AI agents without adding extra compute, saving time and cost on complex optimization problems.
