AgentEvolver Achieves 82% SWE-bench Pro Resolution via Self-Evolution
October 9, 2026
AgentEvolver enables agents to evolve reusable capabilities, such as methods and control flows, during task execution without changing the foundation model. The system reached an 82.08% resolution rate on SWE-bench Pro Public by coordinating persistent planning and recoverable context.
HOW THIS AFFECTS YOU
●
builderYou can build agents that improve their own operational logic and interfaces during runtime.
●
researcherThis provides a framework for studying capability growth in fixed-parameter models through system-level evolution.