Evolver turns agent execution traces into reusable Gene assets: success +4.4 pp, tokens down 24%

Evolver: Turning Agent Exploration into Evolving Capabilities

Haoyang Zhang, Qumeng Sun, Xiao Zhang, Xiang Liu, Zhidong Fan, Junjie Wang

cs.SE, cs.AI

2026-11-05

Evolver distills agent traces into reusable Gene assets injected into later sessions: on 320 paired runs, success rises 4.38 points to 87.50% while token use falls 24.24%.

What problem this solves

Coding agents burn enormous compute on exploration: tool calls, failed attempts, recovery, verification. In this paper's baseline setups, a single task averages about 1.64 million recorded tokens. Almost everything the agent learns along the way dies inside the execution trace; the next task, or even a fresh retry of the same one, starts the exploration over.

Existing reuse routes package workflows as skills, evolve prompts from feedback (GEPA, Agentic Context Engineering), or let the agent rewrite its own code (Godel Agent). Evolver builds on the strategy-gene line of work and takes on the systems question: how experience is stored, plugged into existing harnesses like Claude Code and Codex, and maintained over time. The framing includes a sober admission that accumulated experience does not automatically help, since irrelevant or stale strategies add redundant work.

Method

Evolver leaves model parameters untouched and hangs an experience layer outside the agent, in three parts:

Evaluation uses a paired B-E-G protocol: run the task as-is (Baseline), distill genes from those trajectories including failures (Extraction), then rerun the task with gene guidance in a fresh session (Gene). Experience pools are isolated per benchmark, harness, and model, and selection prefers genes distilled from the same task.

Results

320 paired executions (2 harnesses x 2 models x 4 benchmarks x 20 tasks each):

SettingMetricw/o -> w/
OverallSuccess rate83.13% -> 87.50% (+4.38 pp)
Claude CodeSuccess rate85.00% -> 87.50% (+2.50 pp)
CodexSuccess rate81.25% -> 87.50% (+6.25 pp)
OverallAvg tokens per task1,641.1K -> 1,243.3K (-24.24%)
Both-success subset (265 pairs)Avg tokens1,278.2K -> 1,053.6K (-17.57%)

Nine of 16 configurations improved and seven stayed flat; none dropped. By benchmark, Terminal-Bench 2.1 gained the most (64/80 to 71/80, +8.75 pp), SWE-rebench +5.00, GAIA +3.75, while SWE-bench Multilingual sat at 96.25% in both conditions. Token use fell in all four harness-model setups, by 13.55% to 35.96%, with the largest drop on Codex paired with DeepSeek.

Two case studies show where the savings come from: a Terminal-Bench token-counting task (Claude Code + GPT) went from 2,188.8K to 295.6K tokens (-86.5%) and from 112 to 17 model calls; an RxJava concurrency bug fix (Codex + DeepSeek) cut tokens 58.2% and calls from 86 to 52.

Why it matters

For anyone building agent infrastructure, this is an experience layer that requires no model training: it hooks into existing harnesses via session hooks and MCP, with a small integration footprint. The most useful number may not be +4.38 pp but the -17.57% on tasks the agent already solves. Cutting a fifth off execution cost on tasks you were going to solve anyway is real money at the million-tokens-per-task scale. The improvement is incremental, but every piece lands in deployable engineering.

Limitations

The paper concedes the two heaviest points itself. The 320 pairs form a diagnostic set: task selection 'additionally considers observed outcome improvements and token usage', so the headline numbers do not come from a random sample and selection bias may inflate them. The protocol also measures reuse of historical experience, with selection preferring genes distilled from the same task, so +4.38 pp is closer to remembering how the agent solved this exact task last time than to cross-task generalization or long-horizon online evolution. Token accounting covers only the final executions; the extraction stage's compute is excluded.

A few more discounts: 20 tasks per configuration means a single task swings five points, so per-config deltas sit within small-sample noise; the 9-up-7-flat split says gains depend heavily on the model-harness pairing; and the lab (Infinite Evolution Lab, EvoMap) ships the EvoMap network the system is built around, so the paper reads partly as a product announcement.

Terms

Source

What people are saying

Related papers

All paper explainers