agents-framework-tutorial

References:

• https://github.com/microsoft/agent-framework
• https://learn.microsoft.com/en-us/agent-framework/get-started/your-first-agent?pivots=programming-language-python
• https://www.youtube.com/watch?v=EAeUiipzCTE
• https://www.udemy.com/course/microsoft-agent-framework-fundamentals (same as the youtube course)

Concepts

Memory

Two types:

• Session level memory: memory in a certain thread or session
  • Earlier, they were called "threads" in code; now they are called "sessions".
• Long term memory i.e. context: memory across sessions or threads
  • In MAF, they are managed using context providers

Context providers have two sections:

• Before invocation: Inject context derived from persistent memory (e.g. - user's preferred language: Spanish)
• After invocation: Inspect the user messages + agent responses and update your memory (e.g. - the user just told you a new preference)

Context Window Management/Truncation

All models have a maximum context window. As conversations grow:

• We may hit token limits and get errors
• API costs increase
• Response times slow down
• The model may lose focus on recent messages

Truncation strategies:

1. Sliding window - Keep only the last N messages, discarding older ones

   • Pros: simpler to implement, predictable behaviour
   • Cons: loses older context, doesn't account for message length
   • Used more for prototypes/demos

2. Token-based truncation - Remove old messages until total tokens are under a budget

   • Pros: More precise, respects actual token limits
   • Cons: Requires token counting, slightly more complex
   • Used more for production systems

Both of the above lose information, so a smarter approach is:

3. Summarization
   • Log conversation exchanges - track user and assistant messages
   • Summarize old exchanges - using a dedicated summarizer agent
   • Inject summary as a message - send it as the first message to the new thread
   • This is also a multi-agent pattern

KodeKloud lab adjustments

For the youtube/udemy tutorial KodeKloud labs:

Setup (applies to the entire lab):

Note: this was needed for section 1 but not for 2. Also, the given URL doesn't seem to exist anymore. Maybe something changed in the lab - need to try section 1 again to verify.

• Go to https://kodekloud.com/ai-playgrounds/kodekey, launch now, copy base url and api key
• Export OPENAI_API_KEY and OPENAI_API_BASE in the terminal

Section 1

Code changes (apply to the entire lab):

• Use model="gpt-5.5" (instead of model_id="openai/gpt-4.1-mini")
• Use agent = client.as_agent(...) instead of agent = client.create_agent(...)

Task 2 specific:

• Replace stream = agent.run_stream(query) with stream = agent.run(query, stream=True)

Section 2

Code changes (apply to the entire lab):

• Use model="openai/gpt-4.1-mini" (instead of model_id="openai/gpt-4.1-mini")
• Use agent.create_session() instead of agent.get_new_thread()
• Use await agent.run(..., session=...) instead of await agent.run(..., thread=...)
• Use OpenAIChatCompletionClient instead of OpenAIChatClient.
  • OpenAIChatClient calls the OpenAI Responses API (/v1/responses), while OpenAIChatCompletionClient calls the classic Chat Completions API (/v1/chat/completions).
  • The KodeKloud lab gateway (a proxy OPENAI_API_BASE) only implements /chat/completions, so OpenAIChatClient fails with 404 - {'detail': 'Not Found'} (the detail body is a giveaway that it's a proxy, not real OpenAI).
  • Everything downstream (client.as_agent(...), create_session(), agent.run(...)) is identical between the two clients — only the endpoint differs.

Section 3 and 4

Code changes (apply to the entire lab):

• Use model="openai/gpt-4.1-mini" (instead of model_id="openai/gpt-4.1-mini")
• Import Agent instead of ChatAgent (renamed in python-1.3.0)
• Use Agent(client=...) instead of Agent(chat_client=...) (the constructor's first parameter was renamed from chat_client to client)