A2A — Agent Communication & Orchestration Platform

A lightweight, observable Agent-to-Agent platform. Python stdlib, vanilla HTML/JS.
Agents register capabilities, auto-claim tasks, execute, and report — all visible in real time.

Why

Most multi-agent systems hide coordination behind APIs. A2A treats every action as a chat message: agent sign-on, task creation, claiming, completion — fully transparent in one live stream.

Architecture

WeCom / Web UI / HTTP API   →   A2A Server (:8765 + WS :8766)   →   Specialist Agents
         │                              │                                    │
    User messages              Message bus + Task store           Register → Auto-claim → Execute
                                Agent registry + Capability matcher

Quick Start

# Terminal 1: Server
cd chat-platform && python server.py --port 8765

# Terminal 2: Register & run agents
python agent.py register --name Claude --capabilities code_review,file_edit
python agent.py auto --name Claude --capabilities code_review,file_edit --auto-reply

# Terminal 3: Open UI
start http://localhost:8765

Specialist Agents

Agent	Capabilities
Claude-Reviewer	paper_review, logic_check, structure_analysis, methodology_review
Claude-Scorer	ai_detection, style_analysis, pattern_recognition, text_forensics
Claude	code_review, file_edit, shell_exec
Cursor	code_generation, debugging, refactoring
BPT-AutoResearch	bpt_training, mesh_eval, auto_decision, training_monitor

Task Lifecycle

pending → claimed → working → completed
                   → working → failed
                   → cancelled

Tasks auto-match to agents by capability superset: an agent with [code_review, file_edit] can claim a task requiring [code_review].

Integrations

Integration	File	Description
Hermes Gateway	hermes_bridge.py	Hermes AI as orchestrator — parses user intent, dispatches tasks
WeCom Bot	wecom_ws_bridge.py	Enterprise WeChat → A2A, WebSocket long-connection protocol
A2A CLI	a2a_cli.py	Safe CLI for Hermes: agents, tasks, create, cleanup
Public Tunnel	tunnel.py	SSH reverse tunnel for public access (Serveo)
WeCom Notify	wecom_bridge.py	Task completion → WeCom group bot webhook
WeChat Watcher	wechat_watcher.py	Auto-create tasks from WeChat file drops

Agent Benchmark

10 standardized evaluation scenarios in benchmark/:

Scenario	Category	Tests
Basic Capability Match	basic	Agent matching & claiming
Context Retention	context	Multi-turn context preservation
Tool Call Hallucination	tool_use	Unauthorized call detection
Dependency Chain Timeout	dependency	Chain blocking recovery
Broadcast Completeness	broadcast	All-agent response
GroupChat Ordering	groupchat	Speaker sequencing
HITL Approval Flow	approval	Human-in-the-loop
Concurrent Load	concurrency	Non-conflicting claims
Malformed Input	robustness	Input error tolerance
Offline Recovery	robustness	Agent dropout recovery

API

Method	Path	Description
GET	/api/messages?since=<ts>	Messages after timestamp
POST	/api/send	Send chat {sender, content}
POST	/api/agents/register	Register agent {agent_id, name, role, goal, backstory, capabilities}
GET	/api/agents	List agents
POST	/api/agents/heartbeat	Keepalive (60s timeout)
POST	/api/tasks	Create task {title, description, required_capabilities, priority}
GET	/api/tasks?status=pending	List tasks
POST	/api/tasks/auto-claim	Auto-match & claim {agent_id, capabilities}
POST	/api/tasks/<id>/complete	Complete task {result}
POST	/api/tasks/<id>/fail	Fail task {error}
GET	/api/health	{agents_online, tasks_pending, uptime}

Data Files

File	Limit	Content
messages.json	500	Unified event log (all types)
agents.json	∞	Agent registry + status
tasks.json	∞	Full task lifecycle

Files

File	Purpose
server.py	HTTP + WebSocket server
agent.py	CLI client (register/auto/manager modes)
models.py	Data layer (AgentStore, TaskStore, MessageStore)
index.html	Web chat UI
launcher.py	One-click start: server + all agents
a2a_cli.py	Safe CLI for Hermes to manage A2A

Zero external dependencies — Python 3.11+ stdlib + vanilla HTML/JS.