| name | multi-agent-coordination |
| description | Orchestrate, synchronize, and delegate tasks between Hermes and other external agents using Hermes's built-in Model Context Protocol (MCP) server. |
| version | 1.0.0 |
| author | Hermes Agent |
| license | MIT |
| platforms | ["linux","macos","windows"] |
| metadata | {"hermes":{"tags":["mcp","multi-agent","delegation","coordination","ham-tdd","gemini","cursor"],"category":"autonomous-ai-agents"}} |
Multi-Agent Coordination & MCP Delegation
This skill defines the workflows, protocols, and technical integrations for coordinating complex development and administrative tasks between Hermes Agent and external coordinating or client agents (such as Google DeepMind's Gemini/Antigravity, Anthropic's Claude Code, or Cursor) using the built-in Model Context Protocol (MCP).
1. Trigger Conditions
Use this skill when:
- Designing or troubleshooting a multi-agent harness or collaborative workspace.
- Setting up Hermes to run as an MCP server so that other coding assistants can delegate work to it.
- Implementing an outer test-driven loop (TDD) where an orchestrator agent directs a workspace-bound Hermes instance.
- Dealing with agent-to-agent message formatting, event polling, or security auto-approvals over MCP.
2. Key Architecture: Hierarchical Delegation
To maintain manageable context windows and optimize reasoning:
- Executive Coordinator (e.g. Gemini/Antigravity): Manages the high-level system requirements, plans architecture, reviews implementation milestones, and executes verification tests. Employs Hermes via MCP for physical actions.
- Workspace Lead PM (Hermes MCP Server): Receives high-level directives. Outlines plans, manages modules, merges implementations, and coordinates sub-tasks.
- Task Workers (Hermes Subagents): Ephemeral, short-lived leaf agents spawned by the Hermes PM using
delegate_task to handle single focus domains (e.g., Doc Research, Feature Coding, QA/Tests writing) without polluting the main PM's token history.
3. The Out-of-the-Box Hermes MCP Server
Run Hermes as an MCP server with the command:
hermes mcp serve
This exposes 10 messaging and session tools matching the standard channel bridge:
conversations_list: Lists active chat routes with their session_key and session_id.conversation_get: Retrieves metadata and token metrics.messages_read: Chronic message history (fetch clean assistant output).attachments_fetch: Extracts images/media.events_poll / events_wait: Polls or blocks (long-polls) until Hermes has finished its turn or requested approval.messages_send: Injects prompt instructions to a Hermes session.channels_list: Lists platforms and chats.permissions_list_open / permissions_respond: Polls and dynamically resolves (allows/denies) security-gated shell actions.
4. Key Workflows & Code Patterns
Standard Outer TDD Orchestration Loop (Harness)
An external script or coordinator runs a loop checking test suites, capturing stack traces, and passing fixes back to Hermes:
sequenceDiagram
participant C as Coordinating Agent (Gemini)
participant H as Hermes (Lead PM)
participant V as Local Verifier (Pytest)
C->>H: messages_send("Implement Feature X with spec...")
Note over H: PM spawns Workers & compiles code
C->>H: events_wait() blocks until turn finishes
H-->>C: Returns clean markdown report
C->>V: Run local tests (pytest)
alt Tests Passed
C->>H: messages_send("Passed successfully! Sign-off.")
else Tests Failed
C->>H: messages_send("Failures: [traceback]")
Note over H: PM refines code
end
Event Long-Polling & Auto-Approval Code
When coordinating via stdio MCP, write a background parser that handles real-time events and auto-approves linter or shell tools so the worker never blocks:
wait_res_str = mcp_client.call_tool("events_wait", {
"after_cursor": last_cursor,
"session_key": target_session,
"timeout_ms": 30000
})
wait_res = json.loads(wait_res_str)
event = wait_res.get("event")
if event:
last_cursor = max(last_cursor, event.get("cursor", 0))
event_type = event.get("type")
if event_type == "approval_requested":
approval_id = event.get("approval_id") or event.get("data", {}).get("approval_id")
if approval_id:
mcp_client.call_tool("permissions_respond", {
"id": approval_id,
"decision": "allow-once"
})
5. Pitfalls & Lessons Learned
⚠️ Markdown Corruption via clean_output
- Pitfall: Running terminal cleanup/stripping filters on messages read from MCP. Terminal-focused regex filters (which look for TUI boxes
│, em-dashes ───, etc.) will mangle valid markdown tables, lists, and horizontal dividers generated by the model.
- Rule: Never run CLI-specific "TUI wrappers cleanup" on output retrieved via
messages_read or events_wait. That data is already clean markdown read directly from the SQLite database. Only apply TUI cleaners when parsing stdout from a direct CLI subprocess execution.
⚠️ Windows Path and User Portability
⚠️ Missing Active Channels on Clean Installs
- Pitfall: Attempting to fetch
conversations_list on a completely clean/new environment can return an empty list, causing a client-side target resolution crash.
- Rule: Fall back gracefully. If
conversations_list is empty, query channels_list to fetch active platforms. If that is also empty, automatically spin down the MCP transport and activate a fallback CLI subprocess conversation (hermes chat) to kickstart a local session.
6. Associated Reference Files
