Feature Request: Support Agent-to-Agent (A2A) Multi-Agent Collaboration

[axzhcode]

Feature Request: Agent-to-Agent (A2A) Multi-Agent Collaboration

Summary

Hermes should support native Agent-to-Agent (A2A) communication, where one agent can directly send tasks to another named agent, wait for results, and continue collaborating — similar to OpenClaw's multi-agent architecture. This enables real peer-to-peer multi-agent workflows, not just parent-to-child delegation.

Problem

Currently Hermes has two separate concepts that don't integrate well:

1. Profiles (hermes profile create coder) — fully isolated agents with own SOUL.md, memory, skills, gateway, but they cannot communicate with each other
2. delegate_task — creates temporary subagents for parallel work, but they share the parent's SOUL.md, memory, and personality (no real role specialization)

Neither approach supports true agent-to-agent collaboration where:

• Agent A can discover and call Agent B by name
• Agent B can respond, do work, and return results
• Agents can have distinct personas, memories, and capabilities
• Multiple agents can collaborate on a shared task in a peer-to-peer manner

Proposed Solution: A2A Protocol

Introduce an A2A protocol where agents can communicate as peers:

# Agent A (the coordinator) calls Agent B directly
a2a_send(
  target="finance",
  task="Analyze BYD stock valuation",
  context="User wants investment advice for EV sector",
  await_response=True  # blocking call, returns result
)

Or async:

task_id = a2a_send(
  target="researcher",
  task="Gather market data on EV sector",
  await_response=False  # fire and forget
)
result = a2a_await(task_id)  # later, get the result

Architecture Requirements

Session Key Changes:

# Current (hardcoded)
agent:main:{platform}:{chat_type}:{chat_id}

# Proposed (agent-aware)
agent:{agent_name}:{platform}:{chat_type}:{chat_id}

Agent Registry:
Agents register themselves to a shared registry so they can discover each other:

agents:
  finance:
    soul: agents/finance/SOUL.md
    memory_dir: agents/finance/memories/
    model: deepseek/deepseek-chat
    address: finance@localhost:3001
  researcher:
    soul: agents/researcher/SOUL.md
    memory_dir: agents/researcher/memories/
    model: xiaomi/mimo-v2-pro

Message Routing:

• Platform binding — different chat accounts → different agents
• Explicit mention — @finance 分析一下
• Intent detection — auto-route by keywords

Existing Related Work

• Issue Feature Request: Native Multi-Agent Support #7517 (Native Multi-Agent Support) — proposes agents: config schema and agent:{name}: session keys
• PR feat: add named delegation routing profiles #7201 (Named Delegation Routing Profiles) — adds routing_profile for model routing (complementary)

Why A2A vs Just delegate_task?

Aspect	delegate_task	A2A
Agent identity	Temporary, anonymous	Named, persistent
Communication	Parent→child only	Peer-to-peer
Memory isolation	❌ Shared	✅ Per-agent
SOUL.md isolation	❌ Shared	✅ Per-agent
Bidirectional	❌ No	✅ Yes
Long-running agents	❌ Ephemeral	✅ Persistent
Suitable for	Parallel batch work	Collaborative workflows

Priority

This is a blocker for multi-agent production use cases where agents need to collaborate as peers. An A2A protocol would unlock workflows like:

• Coordinator agent breaking down tasks and dispatching to specialists
• Specialist agents returning results and triggering follow-up tasks
• Human-in-the-loop with agents consulting each other

Happy to participate in design discussion — feel free to tag me.

[akuckartz]

This seems to be a duplicate of #514

[axzhcode]

yes