Skip to content

Hermes agent

The A2A gate hosts one Hermes agent instance (on node-2) to demonstrate that ai-memory's A2A semantics work with a second, independently-implemented agent framework's MCP client.

What Hermes brings to the A2A gate

  • A framework-independent validation axis. If scenarios pass for OpenClaw but fail for Hermes (or vice versa), the regression is framework-specific, not ai-memory-specific — which is exactly what the A2A gate exists to distinguish.
  • A third agent_id value (ai:bob) to round out the minimum- three-agents requirement for scenarios 6 and 7.
  • A different MCP client library than OpenClaw's, which catches subtle schema-compatibility bugs that might be hidden inside a single implementation.

Per-node setup (scripts/setup_hermes_agent.sh — to be written)

1. apt-get update + install runtime deps
2. Install Hermes from release tarball or container image
3. Configure MCP client to target node-4:
     AI_MEMORY_MCP_ENDPOINT=http://10.260.0.14:9077
     AI_MEMORY_AGENT_ID=ai:bob
4. Provision the scenario-runner script
5. Health-check: agent CLI `hermes version` + echo-request to
   ai-memory

What the A2A gate measures against Hermes

Same contract as OpenClaw — tool-invocation correctness, identity preservation, scope honoring — but exercised through Hermes's MCP client to catch any framework-specific divergence.

Where the assertions in scenarios differ between OpenClaw and Hermes (they shouldn't, but if they do), each framework's per-scenario behaviour is recorded separately so divergences are visible.

Framework-to-framework handoff

Scenario 2 (shared-context handoff) specifically tests the flow:

  • Agent A on node-1 (OpenClaw) writes a handoff memory.
  • Agent B on node-2 (Hermes) recalls it.
  • The memory round-trips cleanly between two different MCP client implementations talking to the same ai-memory store.

A red result on this scenario where the ship-gate is green narrows the diagnostic immediately: the regression is in one of the two frameworks' MCP clients, not in ai-memory's server handlers.

Version pinning

Same convention as OpenClaw — hermes_version is recorded in every scenario artifact. Defaults to latest stable; overridable via workflow input.