Limit the conversation context passed to the A2A agent

(1) Share

Report

Posted on by junseongE

I got a question while studying the Microsoft Learn module Build cross-platform multi-agent solutions using the Agent2Agent protocol in Microsoft Copilot Studio.

Q5. Before deploying an A2A integration to production, a maker reviews the message payload and finds the conversation history sent to an external partner's agent includes a customer's account number captured earlier in the same session. The external agent only needs the current shipment query. What should the maker do?

A: Review the external agent's data processing terms and limit the conversation context passed to the A2A agent to only what it needs for the task.

I understand it theoretically, but how can we limit the conversation context passed to the A2A agent actually? Is this possible in Generative orchestration mode?

If not, can/should we trigger the A2A agent explicitly within the topic node without using Generative orchestration mode?

Categories:

General topics

I have the same question (0)

All responses (3)

Answers (0)

Sort by

Prasad-MSFT Microsoft Employee on at

Like
a
(0)

Report
Copy link

Link copied!

You are correct that the Learn module is referring to the data minimization principle. However, in the current Copilot Studio A2A experience, there is no maker-facing setting to explicitly control or trim the conversation history sent to a connected A2A agent.

In Generative Orchestration, the runtime determines the context shared with the external agent.

If strict control over shared data is required, the recommended approach is to use an intermediary flow/API/custom action and pass only the necessary parameters (for example, the shipment query) instead of relying on direct A2A delegation.

Was this reply helpful? Yes No
Suggested answer

11manish 3,563 on at

Like
a
(0)

Report
Copy link

Link copied!

At present, Generative Orchestration does not provide a supported way to selectively limit or redact the conversation history passed to an A2A agent.

If you need precise control over the information shared with an external agent, the recommended approach is to invoke the A2A agent explicitly from a topic (or

another controlled workflow), disable conversation history if possible, and pass only the minimum required inputs.

This approach best aligns with the security guidance described in the Microsoft Learn module while adhering to the principle of least privilege and data

minimization.

Was this reply helpful? Yes No
Suggested answer

Haque 3,797 on at

Like
a
(1)

Report
Copy link

Link copied!
Hi @junseongE,

I was also going trhough taht learning path, what I percieved is the entire context is based on data security in terms of sensitive data or redundant/unnecessary processed by the other agent in a A2A context.

Basically, in a A2A scenario, generative orchestration forwards the full conversation context or session memory by defaul, which can include sensitive or unnecessary data. The simple idea is to create a sanitzed or minumum payload that excludes sensitive information like account numbers.

Answers on your quesion to be simplied:

Q1: How can we limit the conversation context passed to the A2A agent actually? Is this possible in Generative orchestration mode?

A1: No atuo-magic ways are there to it directly. We can implment explicit filtering or parameter passing based on situation in the orcestration logic.

Q2: If not, can/should we trigger the A2A agent explicitly within the topic node without using Generative orchestration mode?

A2: Usaully, explicit calls with needed and defined input params allow us to limit congtext and avoid sending sensitive or unnecessary data

References:

Multi-agent pattern and practice

Orchestrate Agent Behaviour with GenAI

I am sure some clues I tried to give. If these clues help to resolve the issue brought you by here, please don't forget to check the box Does this answer your question? At the same time, I am pretty sure you have liked the response!

Was this reply helpful? Yes No