Evolver paper turns agent exploration into reusable strategy assets

According to an article posted on November 6, 2026 by cnblogs, the EvoMap research team at Infinite Evolution Lab presented Evolver in the paper "Evolver: Turning Agent Exploration into Evolving Capabilities". Evolver is an external experience system for AI agents. It does not train or fine-tune a foundation model. Instead, it turns the procedural knowledge produced while an agent carries out a task into reusable strategy assets that can be retrieved and withdrawn, then injects those assets into the context of later tasks.

The article starts from a recurring problem: after an agent finishes a task, what remains is often more than an answer. The run leaves behind an expensive exploration path, including the files that were read, the tools that were tried, the places where calls failed, and the checks that confirmed the result. Because model parameters do not change and a new session does not naturally inherit the previous round's judgement, that paid exploration cost tends to be paid again on the next run. Saving the full log does not by itself solve this, because a log answers "what happened this time" while a strategy answers "given this situation, how should I act".

Evolver moves the reusable guidance outward in advance and makes the strategy the main unit of reuse, rather than feeding the complete trajectory back into each new task. The system works as a three-step loop: select a strategy, execute the task, then update experience. Each strategy is described with a summary, steps, applicability conditions, matching signals, and a verification plan. The paper contrasts strategies with skills, which it describes as packaged instructions, workflows, and resources, while strategies add applicability conditions, matching signals, and a verification plan. The engineering question the article highlights is how to turn one-off execution records into actionable guidance that has stated conditions, can be retrieved, and can also be withdrawn.