Codex × Claude Code × Hermes: turn three agents into one working system

01 / Blueprint

The right mental model: not three chat boxes, but a three-layer system

Many teams install Codex, Claude Code, Cursor, and Hermes, yet still paste the same context every day. The advanced move is to divide the system into execution, judgment, and operations.

02 / Division of Labor

Task split: do not ask every agent to do the same job

03 / Shared Memory

Let all three tools read the same project operating manual

The practical lesson from official docs and community practice is consistent: advanced agent workflows are not powered by magical prompts. They work because the project has durable context. Codex reads AGENTS.md, Claude Code reads CLAUDE.md, and Hermes uses skills, memory, and session handoff.

your-project/
├─ AGENTS.md          # Codex / general coding-agent rules
├─ CLAUDE.md          # Claude Code interaction and architecture rules
├─ docs/agent/
│  ├─ architecture.md # system architecture and data flow
│  ├─ testing.md      # test, lint, and build commands
│  └─ release.md      # deploy, rollback, and human-approval gates
└─ .hermes/plans/     # optional: Hermes plans / handoffs

04 / Recipes

Three hybrid workflows you can actually run

Please do not edit files yet.
1. Read AGENTS.md / CLAUDE.md / docs/agent/* first.
2. Propose a plan in 3 steps or fewer.
3. List files to change and how to verify them.
4. Wait for confirmation before implementation.

# 1. Install and sign in to Codex
codex

# 2. If your version supports it, start the MCP server
codex mcp-server

# 3. Register it in Claude Code
claude mcp add --transport stdio codex -- codex mcp-server

# 4. Check it inside Claude Code
/mcp

Please turn this repository release process into a Hermes skill:
- trigger conditions
- files to read first
- verification commands to run
- actions requiring human approval
- handoff / report to leave at completion

05 / Templates

Copy-ready templates

# Project Agent Guide

## What this project is
One-sentence product, user, and data-flow summary.

## Commands
- Install: `...`
- Dev: `...`
- Test: `...`
- Build: `...`

## Conventions
- Main language / framework
- File placement rules
- Naming and error-handling principles

## Definition of Done
- Tests and lint pass
- Human-readable summary exists
- Major changes include rollback notes

## Never do
- Do not read or print secrets
- Do not deploy production directly
- Do not run destructive database changes

Please review only; do not edit files.
Check this diff for:
1. compliance with AGENTS.md / CLAUDE.md
2. security, data-loss, permission, or migration risk
3. unverified assumptions
4. missing tests
Classify findings as Critical / Major / Minor.

Use when preparing or reviewing a production release:
- inspect current git diff and deployment config
- verify tests, build, and smoke checks
- identify human-approval gates
- produce release handoff and rollback notes
Do not deploy unless explicitly approved.

## Agent Handoff
- Goal:
- In scope / out of scope:
- Files changed:
- Commands actually run:
- Test / build result:
- Risks:
- Human approval needed:
- Next best action:

06 / Guardrails

The stronger the multi-agent workflow, the more you need hard guardrails

AGENTS.md, CLAUDE.md, and prompts are still text. Real safety comes from environment isolation, permission boundaries, and verification gates.

• Default every agent to a feature branch or git worktree, not direct main edits.
• Codex, Claude Code, and Hermes should not touch production secrets; use variable names, never values.
• Start MCP with read-only documentation tools; require approval before writing external systems.
• Deploys, database migrations, deletions, and external sends require human approval.
• Completion means real tests, build, and smoke checks passed — not that an agent said it is done.
• When a second agent reviews, make it review artifacts and diffs, not the first agent's self-report.