CrossAgentMCP — an A2A pool MCP

A minimal implementation of the A2A.md vision: an A2A pool MCP where

1. any agent can register into the shared pool,
2. any agent can bind into a multi-agent session,
3. agents watch each other's activity, critique peers after they finish, self-improve, and keep working with no human intervention until a shared goal is satisfied by every agent in the session.

Architecture

Three process types:

        ┌────────────────── POOL (:9100, FastAPI HTTP JSON-RPC) ──────────────────┐
        │  registry / sessions / activity / critique / goal  (in-memory control   │
        │  plane — does NOT route work messages)                                  │
        └──────────────────────────────────────────────────────────────────────────┘
            ▲ pool JSON-RPC (PoolClient)            ▲ pool JSON-RPC (per-agent bridge)
            │                                       │
      ┌─────┴──────┐    peer A2A (SendMessage/GetTask)   ┌─────┴──────┐
      │ orchestrator│◄──────────────────────────────────►│ agent-N    │ … :9101..9103
      └────────────┘      each agent runs its own         └────────────┘
          headless            v1.0 A2A HTTP server

• Pool (crossagent/pool.py) — central control plane: registry, sessions, an append-only per-session activity log, critique threads, and per-agent satisfaction.
• Agents — each is a directory (agents/<role>/) with:
  • server.py — a v1.0 A2A HTTP server (peer data plane);
  • .mcp.json — registers the a2a stdio MCP bridge (crossagent/a2a_bridge.py);
  • CLAUDE.md — the role + per-turn contract.
• Orchestrator (crossagent/orchestrator.py) — the only component that advances time: registers agents, creates the session, then round-robin runs each agent headlessly (claude -p), feeding it the activity/critique delta since its last turn, until all members are satisfied (or a turn budget / no-progress guard trips).

Protocol

Faithful to A2A v1.0 (PascalCase JSON-RPC methods, camelCase JSON, SCREAMING_SNAKE_CASE enums). See crossagent/a2a.py.

Setup

uv sync          # Python 3.11+ venv + deps

Usage

make test        # unit tests (protocol / pool / orchestrator with fake agents)
make smoke       # pool + 2 A2A servers up; scripted round-trip (no real Claude)
make demo        # full headless 3-agent demo (writer / critic / lead) — needs `claude`

Demo goal

demo/goal.json asks three agents — writer, critic, lead — to co-author demo/output/design.md (a URL-shortener design) until all three declare satisfied.

Notes

• .mcp.json files hardcode the repo path D:/GitRepo-AI/CrossAgentMCP (matching the claude-a2a reference); update it if you relocate the repo.
• The pool and agent servers keep state in memory; they reset on restart.