Evolver: Turning Agent Exploration into Evolving Capabilities
Haoyang Zhang and colleagues at EvoMap's Infinite Evolution Lab present Evolver, a framework that stores procedural knowledge from agent exploration as structured "Gene" assets and reuses it in later runs of existing harnesses such as Claude Code and Codex, without changing model weights.
Ask this paper
Mechanism. Evolver covers experience acquisition, retrieval and feedback-guided revision of Gene assets, and plugs into existing agent harnesses.
Evaluation. It is tested with Claude Code and Codex, two model configurations, on Terminal-Bench 2.1, SWE-bench Multilingual, SWE-rebench and GAIA.
Results. On a selected diagnostic set of 320 paired runs, task success rises from 83.13% to 87.50% (+4.38 points) and recorded token use falls 24.24%.
Efficiency on solved tasks. Among 265 pairs solved under both conditions, tokens fall 17.57%, so the saving appears even when success does not change.
Caveat. The evaluation set was selected for diagnosis. Transfer to unseen tasks is listed as future work.
Abstract
Agent exploration produces valuable procedural knowledge, yet much of this experience remains confined to transient execution traces and is not systematically reused. We introduce Evolver, an experience-driven framework that transforms agent exploration into persistent, reusable capabilities without updating foundation model parameters. Built around structured Gene assets, Evolver integrates experience acquisition, retrieval, and feedback-guided evolution to support experience-conditioned execution across existing agent harnesses. We evaluate Evolver using Claude Code and Codex, two foundation model configurations, and four benchmarks: Terminal-Bench 2.1, SWE-bench Multilingual, SWE-rebench, and GAIA. On a selected diagnostic set of 320 paired executions, Evolver improves task success from 83.13% to 87.50% (+4.38 percentage points) while reducing recorded task-execution token consumption by 24.24%. Furthermore, among 265 pairs successful under both conditions, token consumption decreases by 17.57%, demonstrating efficiency benefits even when task success is preserved. These findings suggest that persistent procedural experience can improve both task effectiveness and execution efficiency, providing a practical foundation for transforming agent exploration into evolving capabilities.