Advanced: Agent Dev 07 Teaches Agents to Plan Before They Act

Don't hand a big task to an Agent all at once. First state the goal clearly, break it into small tasks, give each small task an owner, and finally put the results together.

Remember These Points

• Goals should be specific: 'plan a three-day itinerary' is better than 'help me travel'.
• Break the goal into subtasks, and assign each one to a dedicated Agent or process.
• Have the model output structured results such as JSON, so the next step can be processed automatically.
• When unexpected input appears, be able to adjust dynamically instead of rigidly following the plan.

Going Deeper

Breaking down tasks. 'Make a three-day itinerary' can be split into booking flights, booking hotels, renting a car, and personalized recommendations. Each part goes to a dedicated Agent, and finally a summarizing Agent assembles the results. If you want to add 'food recommendations' later, just add one more Agent.

Structured output. Have the planning Agent output a JSON document: which subtasks exist and who each one is assigned to. Downstream programs don't need to guess; they can read it directly.

Dispatch and aggregation. A routing Agent receives 'I want to book a hotel', first checks the registry for which Agents exist and what tools each has, then decides whether to hand it to one Agent or bring several into a group to work on it together, and finally summarizes the result for the user.

Event-driven handling. A plan is not fixed: if the user changes requirements midway, or a subtask fails, it should be able to trigger re-planning.

Image from the course repository.

Try It Yourself

Using 'organize a weekend team outing' as the goal, split it into four or five subtasks, and write out the JSON fields you would want the model to output.

Let Your Agent Teach You

Give the text below to your Agent (Claude Code, ChatGPT, and Gemini all work), and it will guide you step by step like a mentor.

References

• Content adapted from Lesson 7 of microsoft/ai-agents-for-beginners (07-planning-design); the original project is under the MIT license.
• For the original and full code: 07-planning-design
• For all courses, see the Agent course roadmap