Implement generative AI and agentic solutions 2.1 Build generative applications by using Foundry
A Microsoft Agent Framework sequential workflow chains a drafting agent, a legal-review…
Study guide skill: Design workflows, tool-augmented flows, and multistep reasoning pipelines
Question B8-016
A Microsoft Agent Framework sequential workflow chains a drafting agent, a legal-review agent, and a translation agent. The translation agent keeps reacting to the original drafting instructions instead of only translating the reviewed text. What should you configure?
Select one answer.
Answer: C
By default, each agent in a sequential workflow consumes the previous agent's full conversation, including its input messages and responses. You can configure agents to consume only the previous agent's response messages, such as with chain_only_agent_responses=True. Concurrent orchestration runs agents in parallel and Magentic is for open-ended tasks with an unknown solution path, so neither keeps the fixed pipeline.
Deep dive
How it works
In a Microsoft Agent Framework sequential orchestration, each agent processes the previous agent's output in order. What each agent receives is controlled separately.
The default. Each agent in the sequence consumes the previous agent's full conversation: both the input messages given to the previous agent and its response messages. In this pipeline, the drafting instructions travel all the way down the chain, which is why the translation agent keeps reacting to them.
The fix. You can configure agents to consume only the previous agent's response messages. In Python, setting chain_only_agent_responses=True on SequentialBuilder applies this to every agent in the sequence. In Go, the equivalent is WithChainOnlyAgentResponses(true). The translator then sees only the legal reviewer's output.
The pipeline stays sequential, which is what the task needs: draft, then review, then translate.
The other options, and when you'd use them
- Concurrent orchestration: Agents execute in parallel, but translation has to wait for the reviewed text.
- Magentic orchestration: It's designed for complex, open-ended tasks where the solution path isn't known in advance. This path is fixed.
- Leave the default: By default each agent gets the previous agent's full conversation, including its input messages. That's the cause of the problem.
Exam tip
Sequential agent sees too much context = chain only agent responses (chain_only_agent_responses=True). Default = full conversation.
References
Last reviewed: