AgentEvolver lets agents self-evolve during task execution with a fixed foundation model, reporting 82.08% resolution on SWE-bench Pro Public
Synopsis
The authors present AgentEvolver, a system that develops capabilities during task execution while keeping the foundation model fixed: eight entity families are exposed to revision through a common versioned lifecycle, a shared Runtime coordinates ongoing work, and persistent planning with recoverable context preserves task direction and supporting evidence; on SWE-bench Pro Public the team reports an 82.08% resolution rate, exceeding its reported baseline without evolution, and six application cases show retained capabilities entering later website, game, and research work while also documenting incomplete objectives and an unsuccessful strategy.
Figure 1: Task-driven capability evolution. An observatory repair motivates complementary changes across eight entity families. Dependencies between branches lead to shared evaluation, collaborative delivery, and retained capabilities. Plan preserves the goal while Runtime coordinates the work. The scene is a conceptual analogy; revised objects and edges describe this illustrative task rather than a robotics experiment.
arXiv · Page 2Interpretation
The system places reusable components (operations, methods, agents, control flow, interfaces, and supporting state) under a common versioned lifecycle so that changes made during task execution can be evaluated and reused. Relative to agent work focused on whether the final task is completed, this separates improvement of a reusable component from success on the final task, requiring the change to connect to evaluation and subsequent use. Supported by the system design itself: eight entity families under a versioned lifecycle plus a shared Runtime, which is constructive system-level evidence.
On SWE-bench Pro Public, the team reports an 82.08% resolution rate with evolution, exceeding its reported baseline without evolution. Provides a quantitative task-outcome comparison indicating an observable difference associated with the evolution mechanism. A reported figure on a single benchmark; the text does not give sample size, repetition counts, or statistical test details.
Six application cases show retained capabilities entering later website, game, and research work, while also documenting incomplete objectives and an unsuccessful strategy. Extends evaluation from single-task outcomes to reuse of capabilities in later work, and presents both successful and unsuccessful cases. Qualitative case-study evidence spanning several application directions, though the text does not state case selection criteria or scale.
Perspective
The work addresses readers studying capability accumulation under a fixed foundation model, in settings where task execution is treated as a process of capability development rather than one-off solving. The versioned lifecycle across eight entity families and the shared Runtime provide an actionable framework for reusing retained capabilities in later website, game, and research work, and a concrete basis for evaluating component improvement separately from final-task success.
The text is summary-level: it does not give the sample size, repetition counts, or statistical tests for SWE-bench Pro Public, nor the selection criteria and scale of the six cases, so the robustness of the 82.08% figure and its gap over the baseline still requires the original details. The authors explicitly note that independent-task transfer and total development cost remain open questions, which readers should watch when assessing long-term usability.