How does the new Function Calling in Semantic Kernel replace the older Planners?

All answers to this question.

The shift to function calling is a move toward efficiency and reliability. The older planners relied on complex, token-heavy prompts to "force" the model to reason through steps. With Semantic Kernel, we now leverage the native capabilities of models like GPT-4o. The kernel provides the model with a list of available functions, and the model returns a structured request for the one it needs. The kernel executes it and sends the result back. This loop continues until the goal is met. It’s much more stable because it uses the model's specialized training for tool use rather than just raw text generation logic, which often went off the rails.

   Answered 2025-11-05 by Brenda Cooper

Does this mean we lose the ability to see the "thought process" of the AI? I liked seeing the reasoning steps the Sequential Planner used to print out.

   Answered 2025-11-10 by Kevin Malone

• You don't lose it! You can still inspect the ChatHistory object or use the built-in hooks to log the "inner monologue" if the model provides it. It’s actually more transparent now because you can see exactly which function the model intended to call and what parameters it generated.

     Commented 2025-11-12 by Thomas Wayne

It also saves a lot of money on tokens. The function calling API is much more concise than the massive multi-step "Chain of Thought" prompts we used to use with planners.

   Answered 2025-11-20 by Rebecca White

• Exactly. Lower latency and lower costs—it's a win-win for anyone running these agents at scale in a business setting.

     Commented 2025-11-22 by Brenda Cooper