Implement generative AI and agentic solutions 2.3 Optimize and operationalize generative AI systems

An insurance claims process must (1) apply fixed eligibility rules that are already…

Study guide skill: Orchestrate multiple models, flows, or hybrid LLM and rules engines

Question B3-012

An insurance claims process must (1) apply fixed eligibility rules that are already written as code, (2) have an LLM agent draft a decision letter, (3) pause for a human adjuster's approval, and (4) resume later, even after a restart. The steps must always run in this order. What should you build with Microsoft Agent Framework?

Select one answer.

A One reasoning model with high reasoning effort so it applies the rules reliably
B A model router deployment that sends rule checks to a small model and letters to a large model
C A single agent whose instructions describe the rules and the required order of steps
D A workflow that connects a function for the rules and an agent for the letter through explicit edges, with human-in-the-loop and checkpointing ✓ Correct answer

Answer: D

Agent Framework workflows connect agents and functions through explicit execution paths, and Microsoft recommends a workflow when a process has well-defined steps and you need explicit control over execution order. Workflows add human-in-the-loop (pause for external input and resume) and checkpoints (save and restore progress), and the guidance is to use a function rather than an agent when code can handle the task, as with fixed rules. A single agent suits open-ended tasks, and model router or reasoning effort only change how a model is chosen or how it reasons, not the order of steps.

Deep dive

How it works

Microsoft Agent Framework gives you two building blocks, and its overview has a simple rule for picking between them:

Use an agent when the task is open-ended or conversational, you need autonomous tool use and planning, or a single LLM call (possibly with tools) suffices.
Use a workflow when the process has well-defined steps, you need explicit control over execution order, or multiple agents or functions must coordinate.

Workflows are functional and graph-based, and connect agents and functions through explicit execution paths (executors and edges). The workflows page lists the durability features this claims process needs:

Human-in-the-loop: pause for external input and resume execution. That's the adjuster's approval.
Checkpoints and resuming: save and restore workflow progress. That's surviving a restart.

The overview also says that if you can write a function to handle a task, do that instead of using an AI agent. The eligibility rules are already code, so they belong in a function executor, while the decision letter (open-ended writing) belongs to an agent executor. The edges fix the order.

The other options, and when you'd use them

One reasoning model with high effort: Reasoning effort controls how much the model thinks before answering. It doesn't guarantee step order, pausing for approval, or resuming after a restart.
Model router: It routes each prompt to the most suitable underlying model. That's model selection, not process control.
A single agent with the rules in its instructions: Agents suit open-ended tasks. Fixed steps in a fixed order call for a workflow, and code-based rules call for a function.

Exam tip

Fixed order + pause + resume = workflow with human-in-the-loop and checkpoints. If code can do it, use a function, not an agent.

References

Last reviewed: 10 Oct 2026

Ready to test yourself? The quiz asks this and every other AI-103 question, with instant feedback. Practice this in a quiz →