Function Calling and Tool Definitions

An AI Agent is an LLM embedded in a stateful control loop with the capability to perceive its environment, invoke external tools, and inspect execution feedback.

The Tool Schema Contract

Language models do not execute functions directly; they emit a structured request containing the tool name and validated arguments. The client runtime runs the tool and returns the observation back to the model.

{
  "type": "function",
  "function": {
    "name": "query_database",
    "description": "Execute a readonly SQL query against the customer inventory table",
    "parameters": {
      "type": "object",
      "properties": {
        "query": { "type": "string", "description": "Valid SQL SELECT statement" }
      },
      "required": ["query"]
    }
  }
}

The ReAct Pattern (Reason + Act)

Modern agents operate via the ReAct framework:

1. Thought: The model analyzes the current goal and past observations.
2. Action: Emits a function call with parameters.
3. Observation: Receives terminal output or API response.
4. Repeat / Final Answer: Continues until the objective is accomplished or error recovery kicks in.

Agent Guardrails

An agent is a loop — think, call a tool, look at the result, repeat. Left unbounded, that loop can spin forever, burning tokens on the same failed step. Guardrails are the simple settings that bound it: a maximum number of iterations and a fixed policy for what to do when a call errors.

What you'll practice

• Bound an Agent Loop — guided · beginner

A command you'll meet

echo max_iterations=10 > agent.yaml

Open any task below to get the full step-by-step walkthrough — every command is explained, and the lab validator checks what you actually built.

What you'll practice

• Design a Planning Loop — challenge · advanced
• Define an Agent Memory Store — practice · intermediate

A command you'll meet

cat ~/agents/planner.json

An agent without a plan is a very slow script. Open the task below for the full walkthrough.