# EP17 Autonomous Channel Brief

- Title: Hermes 團隊怎麼打造：Hermes Agent 與 OpenClaw 的取捨
- Generated: 2026-05-20T16:30:00+08:00
- Source mode: official/project docs first
- Status: research + script candidate, not audio/publish

## Why This Episode

EP16 explained Hermes as an agent operating system. EP17 should answer the next practical question:

> If Hermes is useful, how do we build a team out of it, and when would OpenClaw be a better fit?

The audience should leave with a decision model, not just a feature list.

## Current Source Signals

- Hermes documentation emphasizes self-improving memory, skills, session search, gateway surfaces, MCP, cron, API server, and profile isolation.
- Hermes API Server exposes OpenAI-compatible endpoints and run/health/job surfaces; multiple profiles can run isolated API servers.
- Hermes MCP docs emphasize tool filtering and opt-in parallel tool calls for safe concurrent read-only work.
- OpenClaw documentation emphasizes open-source self-hosting, many channels, many skills, model agnosticism, privacy-first operation, and sandboxed skill execution.
- OpenClaw agent docs present a four-layer model: model, memory, tool, channel.
- OpenClaw security policy frames the product as local-first infrastructure for a trusted operator, not a shared multi-tenant security boundary.

## Teaching Hook

Open with the mistake:

> You do not build an AI team by opening five chat windows. You build it by deciding where tasks enter, where memory lives, which tools each role can touch, and how humans stop bad work before it becomes public.

## Episode Shape

1. The wrong model: five chat windows pretending to be a team.
2. Hermes team pattern: one core discipline, multiple isolated profiles.
3. Role design: intake, research, builder, reviewer, publisher.
4. Memory design: pocket memory, session search, durable vault, skills.
5. Tool design: built-ins, MCP, filtered dangerous tools, parallel read-only work.
6. Runtime design: gateway, API server, cron, health checks, stop buttons.
7. OpenClaw comparison: broader platform/channel ecosystem vs Hermes governed runtime.
8. Decision rule: choose Hermes for governed agent teams; choose OpenClaw for local-first personal assistant platform; combine only with clear trust boundaries.

## Voice Notes

- Spoken source should say `赫米斯` and `開爪`, not force English pronunciation.
- Keep EP11-like conversational cadence: short lines, direct turns, concrete decision points.
- No full render until the EP16 voice identity/hoarseness issue is fixed through a short canary.

## Source Links

- Hermes docs: https://hermes-agent.nousresearch.com/docs/
- Hermes architecture: https://hermes-agent.nousresearch.com/docs/developer-guide/architecture
- Hermes API server: https://hermes-agent.nousresearch.com/docs/user-guide/features/api-server/
- Hermes MCP: https://hermes-agent.nousresearch.com/docs/user-guide/features/mcp
- Hermes memory: https://hermes-agent.nousresearch.com/docs/user-guide/features/memory/
- Hermes skills: https://hermes-agent.nousresearch.com/docs/user-guide/features/skills/
- OpenClaw docs: https://openclawdoc.com/
- OpenClaw agents: https://openclawdoc.com/docs/agents/overview/
- OpenClaw memory: https://openclawdoc.com/docs/agents/memory/
- OpenClaw tools: https://openclawdoc.com/docs/agents/tools/
- OpenClaw security: https://github.com/openclaw/openclaw/blob/main/SECURITY.md