Dev.to AI 🤖 Ai 👁 0 📖 5 min read

Part 28: Agents across vendors: A2A between two platforms

Part 28 · One Agent, Five Platforms Real enterprises won't run one agent platform. Finance might build on one vendor, procurement on another, and the ERP vendor will ship its own agents too. So the agents need to talk t

Part 28: Agents across vendors: A2A between two platforms

Part 28 · One Agent, Five Platforms

Real enterprises won't run one agent platform. Finance might build on one vendor, procurement on another, and the ERP vendor will ship its own agents too. So the agents need to talk to each other, across vendors. That's what A2A, the Agent2Agent protocol from Part 4, is for.

This part puts two of the series' agents on different platforms and walks through one conversation: discovery, authentication, the task lifecycle, and what happens when something fails. It's based on the A2A specification and the Agentic AI Foundation's v1.0 builder's guide, as I read them on 16 November 2026, plus the platform findings from Parts 11 to 25. Like the builds, it's documentation-only.

One A2A conversation between two platforms

The scenario

Two agents, two platforms:

  • The invoice agent from this series, on MuleSoft Agent Fabric (Part 11), where every agent network starts on an A2A trigger.
  • A supplier-communication agent on Azure Foundry (Part 22), which reads supplier emails. Foundry's A2A tool is generally available on protocol 1.0.

A supplier emails: "Why hasn't invoice 4471 been paid?" The supplier agent doesn't own invoices, so it asks the invoice agent over A2A. Later, once a person has approved a fix, the invoice agent asks the supplier agent to send the reply. That second direction is the one Part 12 promised when it moved supplier messages into a separate agent. The mechanics are the same both ways. This part follows the first.

1. Discovery: the agent card

The supplier agent starts by fetching the invoice agent's agent card, a JSON document at /.well-known/agent-card.json. The A2A specification describes it as advertising the agent's "identity, capabilities, skills, service endpoint, and authentication requirements".

For the invoice agent, the card should advertise one skill: explain the status of an invoice exception. Not "post an adjustment", not "approve". An agent from another vendor gets read-only answers, in the same spirit as the read-only stages from Part 9.

Cards can be signed. The AAIF builder's guide explains that v1.0 specifies a JSON Web Signature over a canonicalised copy of the card. It also makes two points worth repeating:

  • "An unverified card is a URL a stranger chose for you." A fake card can copy a trusted one and change only the endpoint.
  • A signature proves the card wasn't changed. It doesn't prove the signer is someone you trust. That's a separate decision.

Part 29 goes through what belongs on a card, and what never should.

2. Authentication between platforms

The card lists the security schemes the agent accepts: API keys, HTTP auth, OAuth 2.0, OpenID Connect or mutual TLS. The specification says a server "MUST reject requests with invalid or missing authentication credentials".

For this pair:

  • The supplier agent calls with a token for its own Entra agent identity (Part 23), not a shared key and not a person's token.
  • MuleSoft's Omni Gateway validates that token and applies A2A policies before the broker sees the request.
  • The invoice agent decides what that identity may ask about, for example only invoices from suppliers the procurement team handles.

If the invoice agent needs more permission mid-task, the task moves to an auth-required state. The builder's guide adds two rules: credentials should arrive "out of band", not in the message where every agent in the chain can read them, and reaching that state grants nothing by itself.

3. The task lifecycle

A request becomes a task with a state. The specification lists them: submitted, working, input-required, auth-required, completed, failed, canceled and rejected.

For the status question:

  1. Submitted, then working: the invoice agent calls its read tools.
  2. Input-required, if needed: "Which invoice? There are two with that number from different suppliers."
  3. Completed, with an answer: "Held for a missing goods receipt; a fix is waiting for approval."

Some answers take longer. If the exception is waiting for a person, the honest answer arrives hours later. A2A has three ways to wait: streaming over Server-Sent Events, polling the task, or push notifications to a webhook. The builder's guide calls streaming "the wrong one for anything that outlives a deploy" and recommends returning immediately and registering a webhook for slow work. Push delivery is at least once, so the receiving side has to handle duplicates.

4. Failure and timeouts

The cross-vendor version of Part 8's named errors:

  • Out of scope: a request to change an invoice gets rejected, not attempted. The invoice agent doesn't have that skill to offer.
  • Unknown or forbidden: the specification says servers shouldn't distinguish "does not exist" from "not authorised", so a caller can't probe for invoice numbers.
  • Too slow: the caller sets its own limit. When it's reached, the supplier agent tells the supplier "we're checking" and hands the case to a person, rather than guessing.
  • Lost connection: the task is durable, so the caller can fetch it again later. The guide warns that an in-memory task store isn't fit for production.

None of this moves the approval gate. Even if every A2A call succeeds, post_invoice_adjustment still needs an approval_id from a person (Part 6).

5. Who supports A2A, and how

From the five builds:

  • MuleSoft: native. Agent networks run on A2A, with an A2A Connector, A2A Bridge and A2A policies in the gateway.
  • Azure: native on protocol 1.0. Foundry agents can call A2A agents and expose an A2A endpoint, though incoming calls are text-only with no streaming.
  • AWS: AgentCore Runtime hosts A2A servers with an agent card, and Gateway can proxy them. The page I read didn't state the protocol version (Part 25).
  • SAP: A2A 0.3.0 through Agent Gateway. Joule calling external agents is text-only and expects a reply within 60 seconds (Part 18).
  • Boomi: no A2A in the pages I read. Agent Control Tower manages third-party agents, so cross-vendor calls would go through an adapter, such as an API in front of the agent (Part 15).

So "supports A2A" covers a wide range. Before pairing two platforms, check three things: the protocol version on both sides, whether long-running tasks are supported, and any reply time limit.

What I'd take from this

Cross-vendor A2A is workable for questions, with read-only skills, each agent's own identity, and verified cards. It's not yet something I'd use to hand another vendor's agent a write. The protocol is ready for that; the trust between organisations isn't.

Would you let an agent from another vendor call yours?

This series describes a reference build on a fictional company. This part is based only on the A2A specification, the AAIF builder's guide and the vendor documentation summarised in Parts 11 to 25, as of 16 November 2026; no hands-on build has been run yet. Views are my own.

📰 Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.