The Third Umpire · Lesson 10 — The Second Look ← Course

The second look — reflection and self-correction.

The best move in the loop is often noticing a result can't be trusted — and not building on it.

Not every step goes cleanly. A tool returns nothing useful, a reading is ambiguous, an early guess turns out wrong. What separates a robust agent from a brittle one is what it does next: charge ahead on a shaky result, or take a second look?

Reflection is the agent examining its own last step — is this reliable enough to build on? — and, when it is not, changing course: retry, reach for a different tool, or revise an earlier call. This is where seeing the outcome pays off: the agent can catch a weak step and try again, where a one-shot answer is simply stuck with it.

Same shaky reading, two agents. Flip between them and watch where the ruling ends up.

Snicko came back unclear — now what?

Doubt the weak step, don't build on it.

Both agents got the same unreliable reading. One treated it as fact and ruled on sand; the other doubted it, refused to conclude, and reached for a better instrument. That extra reflect step — questioning its own last result — is the whole difference between the wrong call and the right one.

Self-correction is not second-guessing everything; it is spotting when a result is weak, a tool failed, or an assumption broke, and adjusting before the mistake compounds. An agent that never looks twice fails the moment a single step misleads it.

Next: drawing stumps →
▸ Go deeper — how agents catch their own mistakes optional

Reflection as an explicit step

Many agents add a deliberate “critique” turn: after acting, the model is prompted to judge its own result before continuing — did that work? is it enough? what is missing? Making reflection its own step, rather than hoping it happens, is a common way to make agents more reliable.

Recovering from tool failures

Tools error, time out, or return junk. A sturdy agent treats a failed call as information — retry once, try an alternative tool, or report that it could not verify — instead of pretending the empty result was an answer.

Revising earlier conclusions

The hardest self-correction is undoing a step you already took. If later evidence contradicts an earlier call, a good agent updates it rather than defending it — the same humility a review demands when fresh footage overturns a first impression.

Knowing what a result is worth

Reflection leans on calibration: sensing that “inconclusive, loud crowd” is weak while “clear mark on the edge” is strong. An agent that treats every observation as equally solid cannot tell when to look again.