Demo 3 — A real agent loop with Ollama + 2 tools

A small console that wires OllamaAgentProvider to Microsoft.Extensions.AI's IChatClient, exposes two tools (calculator, now), and lets the model decide when to call them. Watch the loop step through prompt → model → tool call → result → final answer.

Prerequisites#

Ollama running locally

A tool-capable model pulled, for example:

ollama pull gemma3:4b
# or: llama3.1, qwen2.5, phi4

Run#

cd sessions\session-2\code\demo3-agent-loop
dotnet run "What is sqrt(2024) rounded to 2 decimals?"
dotnet run "What time is it in Tokyo right now?"
$env:OLLAMA_MODEL = "llama3.1"; dotnet run "Compute 12 * (4 + 3) / 2"

Sample output:

📨 user: What is sqrt(2024) rounded to 2 decimals?

   🛠️  tool: calculator(SQRT(2024))
   ✅ result: 44.99...

💬 assistant:
The square root of 2024 is approximately 44.99.

What's happening under the hood#

OllamaAgentProvider.CreateChatClient(...) returns an IChatClient (the Microsoft.Extensions.AI standard interface).
We wrap it with UseFunctionInvocation() — that's the agent loop: it inspects every model response for tool calls, runs them, appends the results to the conversation, and re-calls the model.
The two AIFunctionFactory.Create(...) instances become the tool manifest the model sees.
ChatToolMode.Auto lets the model decide whether (and which) tools to call.

This is the same machinery the OpenClawNet Gateway and DefaultAgentRuntime use — just spread across more files, with security policies, persistence, and approval gates layered on top.

Try it#

Add a third tool that returns weather for a city (mock data is fine).
Set ToolMode = ChatToolMode.None and ask a math question — watch the model hallucinate without help.
Switch to a non-tool-capable model and observe what happens (some models will ignore the tool manifest and answer directly).

Caveat: Some smaller local models are inconsistent with tool calling. If the model never invokes the tool, try a larger one (llama3.1:8b, qwen2.5:7b, gemma3:4b).