Agent Evolution Advisor

A read-only review skill for long-lived OpenClaw agents.

Agent Evolution Advisor helps you inspect a deployed OpenClaw agent without letting it automatically modify itself.

It focuses on controlled evolution:

• review recent runtime behavior
• detect drift across AGENTS.md, SOUL.md, USER.md, and MEMORY.md
• assess memory hygiene and interaction quality
• identify tooling reliability and workflow improvement opportunities
• recommend safe next steps without changing production state by default

Review first, improve second.

Why it exists

Long-lived agents drift.

Rules diverge, memory gets noisy, answers become less actionable, and repeated manual work never becomes workflow. This skill makes those problems visible early in a structured, low-risk way.

What it does

In strict read-only mode, it helps you:

• audit recent agent behavior
• review rule drift
• check memory hygiene
• evaluate answer quality
• find tooling and workflow inefficiencies
• produce concise, structured review reports

What it does not do

By default, it does not:

• modify code
• rewrite rule files
• patch config
• apply production changes
• start unattended self-improvement loops

This is a deliberately conservative skill for human-approved improvement.

Example prompts

• 复盘最近 2 天 Agent 表现
• 看一下最近回答质量
• 审一下规则漂移
• 检查 MEMORY 和 daily memory 分层
• 看一下最近工具和 workflow 有没有优化点

Installation / How to use

Option 1: Install from a release asset

If you want the easiest path, download one of the release assets:

• agent-evolution-advisor.skill
• agent-evolution-advisor.skill.zip

Recommended:

• use .skill if you want the native skill package
• use .skill.zip if you prefer a more generic download format

Then import the skill into your OpenClaw environment using your normal local skill installation flow.

Option 2: Use from source

You can also clone or download this repository and use the extracted skill directory directly:

agent-evolution-advisor/

Place that directory into your local OpenClaw skills area, or use it as a base for your own modified version.

How to trigger it

After installation, invoke the skill with review-style prompts such as:

• 复盘最近 2 天 Agent 表现
• 看一下最近回答质量
• 审一下规则漂移
• 看一下最近工具和 workflow 有没有优化点

What to expect

This is a read-only review skill.

By default, it will:

• inspect recent runtime behavior
• review rule drift, memory hygiene, interaction quality, tooling reliability, and workflow design
• produce structured recommendations

By default, it will not:

• modify code
• patch config
• rewrite rule files
• change production behavior automatically

Included files

• SKILL.md — workflow, scope, and boundaries
• references/finding-examples.md — finding classification guidance
• references/report-format.md — review output structure
• references/usage-examples.md — trigger phrases and templates

Scope

This is a V1 skill focused on read-only review.

The limitation is intentional: it makes the skill safe enough to use in real deployments before adding more automation.

Core principle

No autonomous production changes by default.