The Third Umpire · Lesson 10 — The Investigation Chain ← Course

The investigation chain — the agent loop.

One tool call answers a question. A chain of them, decided on the fly, settles a case.

Last lesson the model reached for a single tool and answered. But a hard call is rarely one question — it is a chain. The umpire asks what it still needs, consults a tool, reads what comes back, and decides whether that settles it or whether to ask again.

That is the agent loop: think (what do I still need?) → act (consult a tool) → observe the result → decide (enough, or go again?) — and round again, until it renders the verdict. The model drives the loop itself.

What makes it an agent, not a single tool call, is judgement. Because it sees the result of each step before picking the next — unlike a one-shot answer, fixed the moment it is written — it can correct course when a reading is ambiguous, and, hardest of all, know when to stop.

Take one hard appeal and turn the loop, step by step, until the umpire has enough to rule.

An appeal comes in
THINK ACT OBSERVE DECIDE not enough — go again enough RULE

Not one call — a chain of them.

An agent does not answer in one shot. It loops — working out what it still needs, calling a tool, reading the result, deciding whether to stop or go again — with every step steered by what the last one showed. Three consultations here, each chosen because of the one before it, then a verdict that rests on all three.

The tools were Lesson 8. The loop is what turns them into an agent: planning, self-correction, and a stopping rule. Give it a goal and the means, and it runs the review itself — which is why the next question is how far you let it run on its own.

Next: the plan →
▸ Go deeper — planning, stopping, and staying in control optional

Think–act–observe, over and over

This loop has a name in the field: a reason-and-act pattern. The model writes a short thought, chooses an action (a tool call), reads the result back into its context, and repeats. Each turn's observation becomes part of what it reasons over next — that is how later steps depend on earlier ones.

Planning and self-correction

A good agent breaks a goal into sub-questions (edge? bat or pad? legal ball?) rather than lunging at an answer. And it revises: a spike on the microphone looked like an edge, but the agent withheld judgement until hot-spot confirmed it. Reacting to evidence, not its first hunch, is the whole point.

Stopping conditions

The hardest part is knowing when to quit. Agents stop on a condition — enough evidence, a step limit, a budget, or a confidence threshold — because a loop with no exit will keep calling tools forever, burning time and money. “I have enough” is a decision, not an accident.

Guardrails and a human in the loop

Running on its own is powerful and risky. Real systems fence the loop: limit which tools it may call, cap the number of turns, and require a person to approve high-stakes or irreversible actions. Knowing when not to let it run unattended is part of building one well.

Cost is the quiet catch

Every turn is another model call and another tool call. A three-step review is cheap; an agent that loops twenty times per question is not. Fewer, better-chosen steps usually beat a longer chain — the same lesson as few-shot and retrieval.