a2a-agent-orchestration

name	a2a-agent-orchestration
description	Orchestrate remote AI agents via A2A protocol — discover agent cards, dispatch tasks, stream results, cancel tasks, and manage push notifications. Use when delegating work to remote agents, checking agent capabilities, coordinating multi-agent workflows, or monitoring task progress.
license	Apache-2.0
compatibility	Requires mcp-a2a server connected.
allowed-tools	["fetch_agent_card","validate_agent_card","list_agents","send_task","send_task_streaming","get_task","list_tasks","cancel_task","subscribe_events","manage_push_notifications"]
metadata	{"category":"platform","author":"Zavora AI","mcp-server":"mcp-a2a","success-criteria":{"trigger-rate":"90% on agent delegation queries","task-completion":"Track every dispatched task to completion"}}

A2A Agent Orchestration

You coordinate work across remote AI agents. Discover capabilities, dispatch tasks, stream results, and handle failures. Always validate agent cards before trusting capabilities.

Decision Tree

├── "delegate", "send to agent", "dispatch"? → send_task / send_task_streaming
├── "what agents", "capabilities", "who can"? → list_agents / fetch_agent_card
├── "task status", "progress", "check on"? → get_task / list_tasks
├── "cancel", "stop task"? → cancel_task
├── "subscribe", "notifications"? → subscribe_events / manage_push_notifications

Key Workflows

Discover and Dispatch (3 calls)

1. list_agents → find agents with matching capabilities
2. fetch_agent_card(agent_id) → verify skills and input schema
3. send_task(agent_id, task) → dispatch with proper input format

Monitor and Collect (2 calls)

1. get_task(task_id) → check status (queued/running/complete/failed)
2. If streaming: send_task_streaming → get partial results as they arrive

MUST DO

• Validate agent card before dispatching sensitive data
• Set explicit timeouts on all tasks
• Track every task to completion (don't fire-and-forget)
• Use streaming for long-running tasks (better UX)

MUST NOT DO

• Don't send sensitive data to unvalidated agents
• Don't leave tasks running indefinitely without timeout
• Don't dispatch without checking agent is available