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

You move an agentic research workflow from a non-reasoning model with detailed…

Study guide skill: Implement model reflection, chain-of-thought evaluations, and self-critique loops

Question B8-027

You move an agentic research workflow from a non-reasoning model with detailed step-by-step prompts to a reasoning model. Answers sometimes stop before the work is really done. Which two prompting changes follow the reasoning-model guidance? (Choose two.)

Select 2 answers.

  • A Raise reasoning_effort as the first fix whenever quality drops
  • B Define what counts as done and how the model should verify its own work ✓ Correct answer
  • C Keep the step-by-step chain-of-thought instructions, because reasoning models need every step spelled out
  • D Ask the model to print its raw reasoning so that you can check it
  • E State the task, the constraints, and the expected output format instead of prescribing every intermediate step ✓ Correct answer

Answer: B, E

Reasoning models work best with a clear goal, firm constraints, and an explicit output contract, and they don't need every intermediate step prescribed. For agentic or research-heavy workflows, define what counts as done and how the model should verify its own work. Treat reasoning_effort as a tuning knob rather than the first fix, and extracting raw reasoning by any method other than the reasoning summary parameter isn't supported.

Deep dive

How it works

Microsoft's prompting guidance for reasoning models differs from what works for non-reasoning models:

  • Reasoning models work best with a clear goal, firm constraints and an explicit output contract.
  • Unlike non-reasoning models, they don't need you to prescribe every intermediate step. State the task, the constraints and the output format you expect.
  • For agentic or research-heavy workflows, define what counts as done and how the model should verify its own work. That addresses answers that stop before the work is really done.
  • Treat reasoning_effort as a tuning knob, not the first thing you reach for when quality drops.

So the migration means rewriting the prompts: drop the step-by-step script, and add a goal, constraints, an output format, a definition of done and a way for the model to verify its work.

The other options, and when you'd use them

  • Raise reasoning_effort first: The guidance says it isn't the first thing to reach for when quality drops.
  • Keep the step-by-step instructions: Reasoning models don't need every intermediate step prescribed.
  • Print raw reasoning: Extracting raw reasoning by any method other than the reasoning summary parameter isn't supported, may violate the Acceptable Use Policy, and can lead to throttling or suspension. Use the reasoning summary parameter.

Exam tip

Reasoning-model prompts: goal + constraints + output contract + definition of done + self-verification. No step-by-step scripts; effort is a later tuning knob.

References

Last reviewed:

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

All Generative AI & agents questions · AI-103 study guide

Thanks for your feedback!