Microsoft's Fork-and-Flush yields 66.0% relative gain for AI agents

Microsoft researchers found that accumulated chat context can keep long-running AI agents focused on similar solutions. Their Fork-and-Flush method, which branches the agent and clears chat context while preserving the workspace, raised normalized average scores by 66.0 percent relative to single runs on the same compute budget.

Sources

• reported The Decoder

Made with AI tools. Spotted a mistake? Tell us