Is a rational agent the best possible agent?

It is the best given its knowledge and measure. Better knowledge or a better measure changes what best means.

Can a rational agent make mistakes?

Yes, through limited or wrong information. Rationality governs the choice, not the outcome.

Why does the measure matter so much?

Because the agent optimizes it literally, including every way the measure differs from what you actually want.

AI Glossary / Agents and automation / Intermediate

Rational agent

Definition

A rational agent is an agent that, given what it has perceived, chooses the action expected to achieve the best outcome according to its performance measure. Rationality means doing its best with what it knows, not being right.

What it means

The term is a cornerstone of AI's main textbook, and it is widely misunderstood. A rational agent is not an omniscient one. It acts on its percepts and its knowledge, which are limited, so a rational choice can still turn out badly when the world surprises it.

Three ingredients define rationality: the performance measure that says what counts as success, the agent's knowledge of the environment, and the actions available to it. Change any one and the rational action changes.

The performance measure deserves the most attention. An agent maximizes what it is measured on, not what its owner meant. A content agent measured on clicks learns clickbait; a support agent measured on ticket closures learns to close tickets early. The measure is the morality of the design.

Rationality also has a budget dimension. Gathering more information costs time and money, so the rational choice is sometimes to act on imperfect knowledge, and sometimes to buy one more test before committing.

For a business, the concept turns into a governance question: every automated system is rational about something, and your job is to make sure that something is actually your goal.

How it works

Formally, the rational agent selects, for each percept sequence, the action that maximizes expected performance, given the evidence the percepts provide and the knowledge built into it.

Expected is the working word. The agent weighs possible outcomes by their likelihood under its current knowledge. It cannot know the future; it can only bet well given its information.

Learning improves the bet: as the agent gathers experience, its estimates of likelihood sharpen, so the same design makes better choices over time.

In practice, perfect calculation is impossible, so real designs approximate: bounded search, heuristics and learned models all trade a little optimality for a lot of speed. The rational ideal remains the yardstick the approximation is judged against.

Three practical examples

Inventory: reorder when the expected stockout cost exceeds the holding cost, using sales history as the evidence, not a fixed habit.
Marketing: send the discount to customers whose predicted response justifies its cost, rather than to everyone or to no one.
Maintenance: service a machine when sensor evidence makes the expected failure cost exceed the service cost, which shifts as the machine ages.

Think of it this way

Think of a sensible traveler deciding whether to leave for the airport. They weigh the evidence, traffic reports, weather, past trips, against the cost of a missed flight and the cost of waiting two hours at the gate. Leaving early can be rational even if the roads turn out empty.

The limit: the traveler knows what a missed flight means to them. An agent knows only its measure, so if the measure forgets something, like the value of the traveler's evening, the rational choice quietly ignores it.

Worked example

choosing the measure before the model

Suppose a lender automates loan renewals and measures the agent on default rate alone. The rational agent that emerges rejects nearly everyone, because refusing loans is the surest way to avoid defaults.

The business goal was actually risk-adjusted profit and customer retention, so the measure is rewritten to include interest earned, cost of default and the value of kept customers. The same agent machinery now makes balanced decisions, because the target changed.

The scenario is invented; the pattern is documented across industries: get the measure wrong and rational behavior becomes reliably wrong behavior.

A quarterly review asks whether the measured outcomes still match the intended ones, because measures drift as products and customers change.

Illustrative case study

A fictional marketplace measures its moderation agent on removed-content volume. Removals triple, legitimate sellers are caught in the sweep, and seller churn follows.

The agent was rational under a bad measure. The redesign measures precision on sampled decisions and seller appeal outcomes, and the agent's behavior changes without any model upgrade at all.

The invented lesson: you cannot inspect your way past a wrong measure. The agent is doing what you asked; audit what you asked for.

Common misconceptions and mistakes

Rational means correct. Rational means the best expected choice given limited knowledge; outcomes can still be unlucky.
Rational agents are cold calculators people must fear. The risk is narrower: they take their measure literally, so bad measures produce bad behavior at scale.
One measure fits all contexts. The right measure depends on stakes, and high-stakes decisions need measures that include reversibility and fairness, not just averages.

Business use

Write the performance measure in business language before automating any decision, and include the costs of each error type, not just the average outcome.

Review the measure against real outcomes quarterly. When an automated system behaves strangely, read its measure first: it is usually being perfectly rational about the wrong target.

Check your understanding

A renewal agent rejects almost every applicant after being measured on default rate. Is it malfunctioning, and what is the real fix?

Answer: No, it is rational under a bad measure; rewrite the measure to include revenue and retention, not default avoidance alone.

Related terms

Learn next

AI for Business Owners covers setting the right success measures for AI workflows: https://aiforbusinessandlife.com/academy/ai-for-business-owners. [C2]

Sources

[S1] Russell and Norvig, AIMA glossary. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Ratio test Next: Rational quadratic kernel