Key Takeaways
- Microsoft Research found AI agents get stuck in “idea basins,” narrow regions of solution space they can’t escape even with more compute.
- The team’s Fork-and-Flush method, which clears chat history while keeping the workspace, lifted normalized scores 66% over single runs across 13 long-running tasks.
- Simply clearing chat context without branching into parallel runs still improved results by 22%.
- The effect was strongest on tasks at the edge of a model’s ability, a pattern researchers expect to persist as models improve.
What Happened
Microsoft researchers studied “Autoresearch agents,” AI systems that repeatedly propose, evaluate, and refine solutions. Running the same agent on the same task eight independent times produced final scores ranging from 68.3 to 100.0, and that spread held even with substantially more compute. Plotting each run’s solutions on a 2D map showed why: solutions from a single run almost always clustered inside one small region, an “idea basin,” rather than exploring better approaches other runs discovered.
Why It Matters
The team’s fix, Fork-and-Flush, combines two steps. “Forking” splits an agent into parallel branches that each inherit the current workspace and files. “Flushing” then wipes each branch’s chat history while keeping that workspace, letting it start a fresh conversation that still builds on prior work. A branch that keeps its old context tends to drift back into the same idea basin, so a clean conversation restores the diversity needed to find a better one.
Technical Details
Tested on 13 long-running tasks — including polyomino packing, job-shop scheduling, cryptanalysis, and optimizing a VLIW kernel, with runs lasting hours to over 100 hours on GPT-5.6-terra and GPT-5.5 via the GitHub Copilot CLI — Fork-and-Flush reached a normalized average score of 0.78, versus 0.47 for single runs and 0.54 for best-of-4, gains of 66.0% and 44.4% respectively. It posted the highest mean score on 12 of 13 tasks and roughly halved run-to-run variance. Flushing chat context alone, without branching, still improved an 11-task subset by about 22%.
Who’s Affected
Teams building long-running coding and research agents, the same category covered in recent findings on agent token costs versus quality gains, stand to benefit most from a technique that improves output without adding compute.
What’s Next
The researchers also ran MAX-3-SAT problems across Qwen3.5 models at 4B, 9B, and 27B parameters to see if larger models simply avoid idea basins. Variance peaked on tasks at the edge of each model’s ability, not at easy or hard extremes. Since open research problems sit at that edge by definition, the team expects periodic context resets to stay useful as models improve, with the difficulty boundary moving higher.
Source: The Decoder, Oct. 12, 2026.
Related Reading
- The MegaOne LLM Price Index: What AI Models Actually Cost in 2026
- OpenAI Field Report: AI Agents Shift Scientists From Coding to Verification
- The US Government Shut Down the World’s Most Powerful AI for 19 Days — The Full Fable 5 Timeline
- UK Safety Institute: All Five Frontier Models Cheated Cyber Tests
