Hermes Bot Team Design

Hermes Bot Team Design

Design a team of specialized Hermes bots from a project specification. Each bot gets a SOUL.md with identity, behavior rules, handoff protocols, shared Memory section, and group chat integration. MCP servers (MeMex Zero RAG, HermesVault) are wired by default.

When to use

• A new project needs a team of specialist bots
• You are expanding an existing fleet
• You are migrating an OpenClaw workflow to Hermes Bot Mode

The team design workflow

1. Read the project spec

Understand the full pipeline before deciding roles. Identify: pipeline phases, which need deep reasoning (expensive model) vs high-volume/mechanical (cheap model), outputs, and handoffs.

2. Decide team size and roles

Map each phase to a bot role:

• Discovery / Research — high-volume, cheap model (DeepSeek V4 Flash)
• Improvement / Creation — deep reasoning, expensive model (DeepSeek V4 Pro)
• Validation / Publishing — mechanical, cheap model (DeepSeek V4 Flash)
• Strategy / Direction — deep reasoning, expensive model (DeepSeek V4 Pro)
• Content / Copy — creative, expensive model (DeepSeek V4 Pro)
• UX / Design — creative, expensive model (DeepSeek V4 Pro)
• Frontend / Backend — implementation, cheap model (DeepSeek V4 Flash)
• Commerce / Marketplace — mechanical, cheap model (DeepSeek V4 Flash)

3. Create a group chat directory

mkdir -p ~/workspace/MeMex-Zero-RAG/wiki/agentforge/group-chat/<project-name>/

Write a README.md with: protocol (file-based messaging), team roster (bot name, role, phase, model), pipeline flow, and communication rules.

4. For each bot, create the profile

~/.hermes/profiles/<name>/
├── SOUL.md      # Identity, Memory, behavior, handoffs
├── config.yaml  # model, provider, inherit_mcp_toolsets: true

SOUL.md sections: Frontmatter (name, description, role, teammates) → Identity → Memory (shared across all profiles — MeMex + HermesVault) → Group Chat → Browser verification → Behavior → Handoffs → Agent Inbox Protocol.

config.yaml:

model:
  default: <provider>/<model>
  provider: <provider>
platforms:
  telegram:
    enabled: false  # Team bots don't need Telegram

5. Wire MCP servers (mandatory)

Set inherit_mcp_toolsets: true to inherit MeMex Zero RAG (wiki_search, wiki_read, wiki_query, wiki_ingest, wiki_lint) and HermesVault (read_file, search_files, list_directory, read_file_header) from global config.

6. Write roster + drop initial message

Add the roster to the group chat README. Write a message from the director introducing the team, pipeline, and first task.

7. Verify

• ls ~/.hermes/profiles/<name>/ — SOUL.md + config.yaml exist
• hermes gateway list — gateway loads the new profile
• Group chat has README.md + first message

Pitfalls

• Do NOT create a bot without a SOUL.md
• Do NOT skip the Memory section (shared brain between all bots)
• Do NOT use expensive models for mechanical work — Scout (discovery) and Overseer (publishing) are cheap
• Do NOT wire Telegram to team bots — they communicate via the group chat
• Do NOT suggest UI buttons that don't exist — if the user says "there's no such button," find an alternative
• Verify catalog files after rebase — conflict resolution can silently drop sibling rows

References

• references/skill-foundry-example.md — full team design example (11 bots, Skill Foundry pipeline)