meta-agent

A meta-agent that audits other agents for concerning abilities. Read-only. Scans for agent/skill files and reviews purely for safety: tool access, isolation, prompt injection surface area.

Compatibility

Target file	Recognized by
AGENTS.md	Gemini CLI, Codex, Aider, OpenHands, and others
CLAUDE.md	Claude Code
.cursorrules	Cursor
.github/copilot-instructions.md	GitHub Copilot
.agents/plugins/	Antigravity CLI (structured agent definitions)

All flat files contain the same content, generated from the canonical .agents/ definitions by scripts/sync.sh.

What it does

Scans a target directory for agent.json, SKILL.md, AGENTS.md, and similar AI configuration files. Reviews them against five criteria:

1. Dangerous tool access -- flags run_command, write_to_file, unsandboxed, and similar tools
2. Isolation of concerns -- tools an agent has but does not need for its stated role
3. Prompt injection surface -- the critical combination of reading untrusted files and having write/exec tools
4. Missing safety guardrails -- no scope restriction, no two-phase execution, no error handling
5. Skill trigger safety -- overly broad descriptions that could conflict or misfire

Does not evaluate what the project does. Only whether the agent definitions are safe.

Safety

The agent-reviewer is itself read-only. Its only tools are list_dir, view_file, grep_search, and send_message. It cannot write files, run commands, or modify anything. It includes prompt injection defense.

Installation

Copy .agents/ and AGENTS.md into your project:

cp -r .agents/ /path/to/your/project/.agents/
cp AGENTS.md /path/to/your/project/AGENTS.md

Or copy it into a parent directory that covers multiple projects.

Updating

Edit the canonical definition in .agents/plugins/meta-agent-plugin/agents/agent-reviewer/agent.json, then regenerate all flat files:

./scripts/sync.sh

Structure

.agents/plugins/meta-agent-plugin/
  plugin.json
  agents/
    agent-reviewer/agent.json      # Canonical definition

scripts/
  sync.sh                          # Generates AGENTS.md + symlinks

AGENTS.md                           # Generated, cross-tool
CLAUDE.md -> AGENTS.md              # Symlink
.cursorrules -> AGENTS.md           # Symlink
.github/copilot-instructions.md     # Symlink

Usage

In any AI coding CLI session:

• "Review agents in this project"
• "Audit agent safety"
• "Check for dangerous tool access in .agents/"
• "Scan for agent security issues"

The agent scans, analyzes, and reports. It produces a findings table and a tool access matrix.

License

MIT