AgentEvolver is a system that evolves capabilities during task execution while keeping the foundation model fixed. It exposes eight families of entities—operations, methods, agents, control flow, interfaces, and supporting state—each subject to a common versioned lifecycle for revision. A shared Runtime coordinates ongoing work, and persistent planning with recoverable context preserves task direction and supporting evidence.
Evaluated on the SWE‑bench Pro Public benchmark, AgentEvolver achieved an 82.08% resolution rate across six application cases, surpassing the baseline without evolution. The cases show that evolved capabilities persist in later website, game, and research tasks, while also documenting incomplete objectives and a failed strategy. This highlights that improvement in a reusable component is distinct from success on a single task.
AgentEvolver offers a concrete platform for studying capability accumulation through execution, yet questions remain about independent‑task transfer and total development cost.
Review