a2a_subagent

OpenClaw A2A Sub-Agent Example

This example starts an OpenClaw runtime in-process, enables the native /a2a surface, and then connects back to it through agent/a2aagent.

It demonstrates a realistic deployment shape:

• a sandboxed OpenClaw runtime hosts bundled skills and host binaries
• another agent process talks to it as an A2A sub-agent
• follow-up turns reuse the same A2A session_id

The example uses the bundled weather skill, so curl must be available on the host.

Run

From the repo root:

cd openclaw

# Export your usual OpenAI-compatible credentials first if needed.
go run ./examples/a2a_subagent \
  -question "What's the weather in Shanghai today?" \
  -follow-up "What about tomorrow?"

If you want to override the default environment-based model settings:

cd openclaw
go run ./examples/a2a_subagent \
  -model gpt-5 \
  -openai-base-url https://api.openai.com/v1

What this example does:

1. starts OpenClaw on a random loopback port
2. enables the native A2A surface at /a2a
3. exposes only the bundled weather skill in the sandbox
4. sends two turns through an A2AAgent
5. shows that the follow-up question reuses the same session

Typical output looks like:

A2A URL: http://127.0.0.1:41237/a2a
Agent: openclaw-sandbox
Published skills: 1

Q1: What's the weather in Shanghai today?
A1: ...

Q2: What about tomorrow?
A2: ...

Notes

• The example intentionally does not use a mock model.
• The remote A2A card is stable by default and does not enumerate every tool. Pass -advertise-tools if you need that behavior.