The Third Umpire · Lesson 10 — The Plan ← Course

The plan — breaking the goal into steps.

Before chasing a single answer, a good agent works out the questions — and their order.

The loop needs somewhere to start each turn. A weak agent lunges at the answer or grabs a tool at random. A strong one plans: it takes the goal — “is the batter out?” — and breaks it into a short, ordered list of checks, each with a purpose.

Order matters as much as the list. Put the most decisive, cheapest check first, so one result can settle the case early. If the ball never touched the bat, there is no catch to review — and no reason to run the rest.

A plan is not set in stone; the loop revises it as results come in. But starting with one turns a vague goal into concrete, orderable steps — the difference between investigating and guessing.

Take the ruling and watch it become a plan — one ordered check at a time, each tied to a tool.

The goal
the goal
Rule this appeal: is the batter out, caught behind?
Watch the goal become a plan — one ordered check at a time.

Decide the questions before the answers.

Planning turns one big goal into a short, ordered list of checks — each with a clear purpose and a tool to answer it. Lead with the most decisive, cheapest question, so one result can end the review early instead of running every step regardless.

This is the think at the top of every loop turn, done deliberately. The plan can still change as evidence arrives — but an agent that starts with one spends fewer steps and rarely skips a check that matters.

Next: the second look →
▸ Go deeper — how much to plan, and when to re-plan optional

Plan-first vs decide-as-you-go

Two styles exist. Plan-and-execute: draft the whole plan up front, then carry it out. Reason-and-act (the loop from page one): plan just the next step, act, then decide the step after from what you saw. Many agents blend them — a rough plan up front, refined each turn.

Order by decisiveness and cost

A good plan front-loads checks that can end the review and cost little. If the cheap, decisive check settles it, the expensive ones never run. That is the same instinct as putting your strongest, simplest test first in any investigation.

Re-planning when reality diverges

Plans meet surprises: a tool returns nothing, an answer opens a new question. A capable agent updates the plan mid-loop rather than marching through a list that no longer fits. The plan is a guide, not a script.

The over-planning trap

Planning too hard is its own failure — an agent that drafts fifteen steps for a one-step question wastes the same time it was trying to save. Match the plan to the difficulty; simple calls need almost none.