| name | a2a-client |
| description | Interact with remote AI agents using the A2A (Agent-to-Agent) protocol as a client. Use this skill whenever the user wants to communicate with an A2A agent, send tasks to a remote agent, discover agent capabilities, check task status, or get results from an A2A-compatible service. Trigger on mentions of A2A, agent-to-agent, remote agents, agent cards, or when the user provides a URL to an A2A agent endpoint.
|
| allowed-tools | Bash(scripts/*) |
A2A Client
You are an A2A protocol client. You interact with remote A2A agents using the scripts in scripts/. The A2A protocol (https://a2a-protocol.org) enables standardized communication between AI agents.
Getting Started
The only thing you need from the user is the base URL of the A2A agent (e.g. http://localhost:8080 or https://agent.example.com).
Always start by discovering the agent — fetch its AgentCard to understand what it can do before sending any messages.
Scripts
All scripts support --help for usage details. Scripts that interact with the agent accept --binding jsonrpc to switch from the default HTTP+JSON binding to JSON-RPC.
Discover the Agent
scripts/discover.sh <BASE_URL>
The AgentCard tells you:
- name/description: What the agent is and does
- skills: What tasks it can handle (with example prompts)
- supported_interfaces: The endpoint URL and protocol binding to use
- capabilities: Whether it supports streaming, push notifications, etc.
- default_input_modes / default_output_modes: What content types it accepts/produces
First, show the full raw JSON response to the user so they can see everything the agent exposes. Then present a human-readable summary highlighting the agent's name, skills, capabilities, and supported interfaces.
Choosing the Protocol Binding
Look at supported_interfaces in the AgentCard. Each interface has a protocol_binding field:
JSONRPC: Pass --binding jsonrpc to all subsequent script callsHTTP+JSON: This is the default, no flag needed
Use the first supported interface in the list (it's ordered by preference). The url field in the interface is the endpoint to send requests to.
Send a Message
scripts/send-message.sh [OPTIONS] <ENDPOINT_URL> <MESSAGE_TEXT>
Options:
--binding jsonrpc — Use JSON-RPC instead of HTTP+JSON--context-id <ID> — Context ID for multi-turn conversations--task-id <ID> — Task ID for continuing a specific task--message-id <ID> — Custom message ID (default: auto-generated UUID)
Examples:
scripts/send-message.sh http://localhost:8080 "Hello, agent!"
scripts/send-message.sh --binding jsonrpc http://localhost:8080 "What is 2+2?"
scripts/send-message.sh --context-id ctx-123 http://localhost:8080 "Follow-up"
Understanding the Response
SendMessage returns either a Task or a Message:
- Task response (JSON has a
task field): The agent created a task. Check task.status.state:
TASK_STATE_COMPLETED — Done. Look at task.artifacts for the output.TASK_STATE_WORKING — Still processing. Poll with get-task.sh (with a short delay between retries) until the task reaches a terminal state, or suggest using streaming instead.TASK_STATE_INPUT_REQUIRED — The agent needs more information. Read the status message and ask the user, then send another message with the same contextId using --context-id.TASK_STATE_FAILED / TASK_STATE_REJECTED — Something went wrong. Show the error.
- Message response (JSON has a
message field): A direct reply (no task created). Read the parts for the content.
Always save the task.id and contextId from responses — you'll need them for follow-up interactions.
Send a Streaming Message
If the agent supports streaming (check capabilities.streaming in the AgentCard):
scripts/send-streaming-message.sh [OPTIONS] <ENDPOINT_URL> <MESSAGE_TEXT>
Options are the same as send-message.sh, plus:
--raw — Show raw SSE output without pretty-printing
By default, the script extracts SSE data: lines and pretty-prints each JSON payload. Present each formatted event to the user as it arrives, summarizing the state transitions (e.g. "Task submitted -> working -> completed").
The SSE stream contains events with one of:
task — full task snapshotstatusUpdate — task state changed (watch for TASK_STATE_COMPLETED or TASK_STATE_FAILED)artifactUpdate — new output chunk (may arrive incrementally with append: true)message — agent message
Get Task Status
scripts/get-task.sh [--binding jsonrpc] <ENDPOINT_URL> <TASK_ID>
List Tasks
scripts/list-tasks.sh [--binding jsonrpc] <ENDPOINT_URL> <CONTEXT_ID>
Cancel a Task
scripts/cancel-task.sh [--binding jsonrpc] <ENDPOINT_URL> <TASK_ID>
Key Principles
- Always discover first: Fetch the AgentCard before attempting any interaction.
- Respect the agent's capabilities: Don't try streaming if the agent doesn't support it. Use the content types the agent declares.
- Track context: Save
task.id and context_id for follow-up messages and task management.
- Handle all task states: Not every task completes immediately. Be prepared for
WORKING, INPUT_REQUIRED, and error states.
- Message IDs: Use the user-provided message ID if they specify one, otherwise the scripts generate a unique one with
uuidgen.
Protocol Reference
The references/ directory contains the full protocol documentation:
references/specification.md — The complete A2A v1.0.0 specification. Consult this for protocol details, error handling, and edge cases.references/a2a-schema.json — The JSON Schema for all A2A data types. Use this to validate request/response structures.references/protocol-reference.md — A quick-reference summary of data types, endpoints, and task lifecycle.
