Choosing an agent framework: three notebooks, three success tests
Start here
In this lesson
- Start with what the framework manages
- Locate the line that supplies New York
- Read Maui against its own request
- Check what kind of output the chart request produces
Start with what the framework manages
Once an application goes beyond one model call, instructions, tools and messages have to work together. The recording introduces Semantic Kernel for production integrations, AutoGen for agent experiments and the Azure service for running a hosted agent. These are the product names used on screen. Consult current SDK guidance before setting up a new project.
Define the deliverable before choosing a notebook
The second notebook introduces an assistant connected to a model client. Do not assume it inherits the first notebook’s destination plugin. Write down the output promised to the user. Two fluent itineraries alone cannot settle a framework decision.
The AutoGen assistant definition 05:54 · Microsoft Learn
Three steps at the screen
- Must the city come from the catalogue?
- Is an unconstrained vacation suggestion sufficient?
- Must the reader receive a chart they can open?
Locate the line that supplies New York
Expand the random-destination function in the Semantic Kernel notebook. Its result is New York. The plan underneath uses that city. If a team must recommend only places it serves, reading the itinerary is insufficient. Compare the list attached to the function, its return value and the destination named in the final reply.
The function call returns New York. 05:20 · Microsoft Learn
Read Maui against its own request
The AutoGen notebook defines a model client and an assistant with no custom tool attached for this run. The sunny-vacation request becomes seven days in Maui. Draw two columns: what did the traveller ask for, and what external data was checked? You can fill in the first. The recording provides no evidence for writing “catalogue verified” in the second.
Seven sunny days in Maui 06:29 · Microsoft Learn
The Maui prompt asks for a sunny vacation. Seven days is the assistant’s proposed duration, not a constraint supplied by the traveller.
Check what kind of output the chart request produces
The Azure example keeps messages in a thread. During the run, Code Interpreter executes Python. The task asks for traveller counts by destination as a chart, so check that a chart appears. Then inspect what its axes and values represent. Producing an image alone does not validate the counts. One agent is enough for this exercise.
The generated chart appears in the notebook. 07:57 · Microsoft Learn
Check the chart against the input
The request mentions operating profit but supplies traveller counts: Bali 100, Paris 356, London 900 and Tokyo 850. People are the unit, so these inputs cannot validate a profit chart. Predict the two tallest bars before looking: London should narrowly exceed Tokyo.
Four traveller counts in the chart request 07:10 · Microsoft Learn
Worked example · Read the run at three stops
Pause | Where to look | How to interpret it |
|---|---|---|
Approved destination | Open the tool call. Compare New York in its result with the final plan. If the values differ, the recommendation fails even if the function worked. | |
Sunny vacation | Maui and a seven-day plan can satisfy the request. No catalogue tool is attached here. An “approved holiday” label would require additional evidence. | |
Traveller distribution | Check that a chart exists and compares cities. If the source of the counts also matters, the image alone cannot establish it. |
Try it yourself · Write an acceptance test
Suppose a travel agency can recommend only partner cities. Which notebook comes closest to providing evidence for this requirement? Write your choice before reading the result. Then reverse the check: the tool returns an approved city, but the assistant writes about a different one. Would your test catch it?
Check your reasoning
The New York example exposes the list and function result. Compare the city in the final plan too. If the next team requests a report, the test changes: open the output and check the variables displayed in the chart.
Troubleshooting
Observation | Check | Expected result |
|---|---|---|
Catalogue compliance | Compare the function result and the final itinerary. | Both should name New York. |
Chart describes the wrong measure | Inspect the input units and chart label. | Passenger counts cannot establish a monetary result. |
Two review questions
- Where would you look for compliance with a catalogue?
- Do the Maui and New York outputs establish a winning framework?
Check your reasoning
- Inspect the call that returned New York, then the list behind it. The city name in a reply is insufficient.
- They use different requests and tool settings. A fair test needs the same job and success criterion.
Short glossary
Model client: the application’s connection to the model. Thread: related messages. Run: an execution of the agent’s task.
Sources and version note
Microsoft publishes Agent Framework migration guidance for Semantic Kernel and AutoGen. The recordings demonstrate behaviours; check current setup instructions for a new project. Microsoft