When another company's agent calls yours: five rules for the handoff
SAP now lets third-party agents invoke its agents through a governed gateway. Here is how to design what your agent accepts, returns and refuses.
On 27 October 2026 SAP announced that it will "allow third-party agents to securely invoke Joule Agents through a governed agent gateway," using open protocols "including Agent2Agent (A2A) and Model Context Protocol (MCP)." Once agents from different companies can call each other, a handoff stops being a screen and becomes a contract. This article is a vendor-neutral way to write that contract before you open the door.
A call from outside is a new kind of intake
Until now, work reached your operation through people and forms: someone wrote, someone filled in a field. An agent calling your agent skips both. It arrives fast, in volume, and without the context a person would add by habit.
A gateway handles identity and access. It does not decide what your agent should do with a request. That part belongs to the operation, and it has five pieces.
- 01Declarewhat the agent offers
- 02Validateshape and limits
- 03Actinside its procedure
- 04Returnresult or reason
- 05Recordone traceable log
Anything that fails validation goes to a named person, not into a retry loop.
1. Publish a short menu, not a capability
List the three or four requests your agent accepts, each with required inputs and a plain description of the result. Everything else is out of scope. A menu is easier to secure, test and explain than "it can do anything the system can do."
2. Validate before acting
Check every inbound request against the menu: known type, required fields present, values inside limits (amounts, dates, quantities). A request that does not fit is rejected with a reason, not interpreted generously. Interpretation is where an outside caller can steer your agent.
3. Keep the agent inside its own procedure
The agent that receives the call runs the same approved steps it would run for an internal request. The caller supplies inputs, never instructions. Text in a request is data. If a field says "ignore your rules", the rules still apply.
4. Return a result or a reason, never silence
Every call ends in one of three states: done, refused with a reason, or handed to a person. Define the third explicitly. Agents that cannot finish tend to retry, and two agents retrying against each other is a loop nobody owns.
5. Give every call one record
Log caller, request type, decision and outcome under a single identifier that both sides can quote. When something goes wrong at 3 a.m., the question "which call?" must have a one-line answer. SAP's release also mentions "observability and administrative controls" for agent behavior, model usage and spend; whatever platform you use, keep the business record yours, not only the vendor's.
Test the contract before you trust it
- Send ten valid requests and confirm the results match what a person would have done.
- Send ten malformed or out-of-scope requests and confirm every one is refused with a reason.
- Send one request with hostile text in a field and confirm the procedure did not change.
- Break the downstream system on purpose and confirm the call ends in a handoff to a named person.
Start with one inbound request type, the least risky one. If you cannot name who receives the refusals, the operation is not ready for the caller.