Microsoft researchers got better agent results by wiping the chat history but keeping the files. Long-running automations can borrow that

An agent's chat history is not just memory. It can also be the thing keeping it stuck.

What happened

Microsoft Research studied long-running agents that repeatedly propose a solution, test it and refine it. Eight independent runs of the same agent on the same task plateaued at very different levels, with final scores from 68.3 to 100.0, and the gaps held even with much more compute. Each run stayed inside a small region of similar solutions, which the authors call an idea basin. Compressing the conversation history periodically did not get them out.

Their method, Fork-and-Flush, splits the agent into parallel branches. Each branch keeps the workspace, meaning the code, files and saved results, but starts with an empty chat context. Without the old reasoning steering it, a branch explores new approaches while still building on prior work.

On 13 long tasks, including job-shop scheduling, packing problems and kernel optimization, with runs from several hours to over 100 hours, Fork-and-Flush scored a normalized average of 0.78 versus 0.47 for single runs and 0.54 for best of 4 on the same compute budget, gains of 66 and 44.4 percent. It had the top mean score on 12 of 13 tasks and more consistent results. Clearing the chat context alone, without branching, improved results by 22 percent. The paper is a preprint on arXiv.

My take

These were research tasks with automatic scores, not CRM workflows. But the design idea transfers well to business automation: keep the real state outside the chat.

How I apply it:

• Save progress to real records, such as a sheet, a database row or CRM fields, after every meaningful step.
• Let a long job restart from those records with a fresh session instead of dragging a huge conversation along.
• When a job has a clear score, like a match rate or a validation check, run a few fresh attempts in parallel and keep the best.
• Treat a long chat history as a cost and a risk, not a free memory.

Agents that can resume from saved work are cheaper to run and easier to debug when something goes wrong.