LLMs & Generative AI

Tool Calling & Agents

Tool calling lets a model request typed external operations. An agent is an application loop that lets the model plan, call tools, observe results, update state, and continue until done or stopped.

Tools are typed capabilities exposed by your application
The model should not own authorization
Agents are control loops
Long-horizon agents need budgets
Agent evals measure task completion

Agent loop sketch

state = initial_task
while not state.done and budget.remaining():
    step = model.plan(state, available_tools)
    if step.kind == "tool_call":
        args = validate_schema(step.args)
        authorize(user, step.tool, args)
        result = run_tool(step.tool, args)
        state = state.with_observation(result)
    else:
        return validate_answer(step.answer)

The model proposes; the application enforces.

Tool safety checklist
Concern	Application responsibility
Schema	Validate types, required fields, ranges
Authorization	Check user/session/tenant permissions
Side effects	Confirm risky actions, make idempotent
Observability	Log tool name, args summary, result, cost
Failure	Handle timeouts, retries, partial results