Agentic Tool-Use Loops

An agent is an LLM placed in a loop where it can call tools (functions, APIs, shell) and act on the results — turning a text predictor into something that does things.

#The core loop (ReAct)

1. THOUGHT  — model reasons about the goal
2. ACTION   — model emits a tool call (name + args)
3. OBSERVE  — runtime executes the tool, returns the result
4. repeat   — until the model emits a final answer

while not done:
    msg = llm(messages)                 # model decides
    if msg.tool_calls:
        result = run_tool(msg.tool_calls[0])
        messages.append(tool_result(result))   # feed back
    else:
        done = True                     # final answer

#What makes or breaks an agent

• Tool schemas — clear names, typed args, good descriptions (Tool Calling & Structured Outputs with Local Models).
• Stopping — cap iterations; agents loop forever without a budget.
• Error handling — feed failures back as observations so the model can recover.
• Context growth — each step adds tokens; summarize or prune (Ollama Parameters Guide).

The local-model catch: smaller models are worse at multi-step tool use. Keep tasks short, schemas tight, and verify outputs. Build the tools as MCP servers.

Related: Building MCP Tools for a Local LLM · Tool Calling & Structured Outputs with Local Models · RAG with Local Embeddings