Agents Talking to Agents

Velo Supply's Juno needs to book a courier. The courier company has its own agent, built by another team, using other tools. Juno can't see inside it, and doesn't need to. They just need a common language, in the way that two companies can exchange an invoice.

• Juno needs a courier
• The courier has its own agent
• They need a common language

MCP and A2A

Here's how the two standards fit together. MCP connects an agent to tools and data. It's vertical. A two A, or Agent to Agent, connects one agent to another. It's horizontal. Google introduced it in twenty twenty five, and it now lives under the Linux Foundation, alongside MCP. They're complements, not rivals. A tool does exactly what it's told. An agent is given a goal, and decides how to reach it.

The agent card

How does one agent find out what another can do? Every A two A agent publishes an agent card, a small JSON document, usually at a well known web address. It gives a name, a description, and a list of skills. It's like a business card for software. Check the spec for the exact fields, as they've changed between versions.

{
  "name": "Velo Supply Returns Agent",
  "description": "Handles returns and refunds for Velo Supply orders.",
  "version": "1.0.0",
  "capabilities": {"streaming": true},
  "skills": [
    {
      "id": "start-return",
      "name": "Start a return",
      "description": "Checks an order and creates a return label."
    }
  ]
}

Long-running tasks

Delegated work can take minutes, or days. So A two A is built around tasks, which move through states. Submitted, working, maybe waiting for more input, and then completed. Here's a small model of that. Run it. The task needed a photo along the way, and it paused, just like our human in the loop graph did. Same idea, across company boundaries.

STATES = ["submitted", "working", "input-required", "working", "completed"]

class Task:
    def __init__(self, task_id):
        self.id, self.history = task_id, []

    def move(self, state, note=""):
        self.history.append(state)
        print(f"{self.id}: {state:<16} {note}")

t = Task("return-881")
notes = ["received", "checking the order", "needs a photo of the damage",
         "photo received", "label created"]
for state, note in zip(STATES, notes):
    t.move(state, note)

return-881: submitted        received
return-881: working          checking the order
return-881: input-required   needs a photo of the damage
return-881: working          photo received
return-881: completed        label created

Summary