How is this different from a simple reflex agent?

A reflex agent acts on the current percept only; a model-based agent acts on an accumulated picture of the world.

Is conversation memory an example?

Yes: the running context a chatbot keeps is a simple world model of the conversation.

What is the classic failure?

Acting confidently on a stale or incomplete internal picture.

AI Glossary / Agents and automation / Intermediate

Model-based agent

Definition

A model-based agent is an agent that keeps an internal model of how its world works, so it can act sensibly even when its current view is incomplete. The model tracks what it cannot directly see.

What it means

A simple reflex agent responds only to what it perceives right now. That fails whenever the right action depends on something no longer visible: the car that just disappeared into the blind spot, the invoice that was already paid yesterday.

A model-based agent maintains an internal picture of the world's state and updates it as new percepts arrive. The model encodes two kinds of knowledge: how the world evolves on its own, and how the agent's actions change it.

The word model here means the agent's working picture of the world, not a machine-learning model, although learned components can help build it. A robot's map, a scheduler's calendar view and a support tool's record of the conversation so far are all world models in this sense.

The design's strength and weakness are the same: the agent acts on its picture of the world, not the world. When the picture is stale or wrong, actions are confidently misdirected. Freshness and correction of the model are the reliability work.

For a business, the question to ask any agent is simple: what does it remember between steps, and how does it find out it was wrong?

How it works

The agent cycle gains a step. After each percept, the agent updates its internal state: given the old state, the action taken and the new percept, what is the world probably like now? Then it chooses an action based on that estimated state, not just the raw percept.

The update rule encodes the model: what persists, what decays, what an action changes. In a delivery robot, the rule is a map and motion physics. In a business agent, it might be which orders are open, which emails were answered and which approvals are pending.

Partial visibility is the design's reason to exist. Because the state accumulates history, the agent can handle situations where the current percept alone is ambiguous, like distinguishing a new complaint from a follow-up.

State can be symbolic records, probabilities, or the running context of a language model. Whatever the form, someone must decide what the agent is allowed to assume when its state is uncertain.

Three practical examples

Support: the agent remembers that the customer already tried the standard fix, so it does not suggest it again.
Scheduling: the agent tracks tentative holds as well as confirmed bookings, so it does not double-book while confirmations are pending.
Warehouse robot: the robot remembers the pallet it saw behind a shelf and drives as if it is still there, because it has no reason to believe otherwise.

Think of it this way

Think of an air-traffic controller tracking aircraft on a board. The controller cannot see the sky; the board is the world for decision purposes, and keeping the board faithful is most of the job.

The limit: the controller's board is refreshed by radar and radio discipline. A software agent's board is refreshed only by the inputs it is given, so gaps in those inputs become gaps in its world.

Worked example

remembering pending approvals

Suppose an agent processes discount requests. A simple reflex version reads each request and acts on the spot, so it grants a second discount to a customer whose first request is still pending approval.

A model-based version keeps a state record: open requests per customer, their age and their status. The second request is recognized as a duplicate of a pending case and merged, not granted.

The state also drives politeness: when the pending approval passes its promised date, the agent drafts a nudge to the approver with the case attached. The scenario is invented; the state-tracking behavior is the pattern.

The design question is what the state must contain for the agent's real decisions, nothing more. Every unnecessary field is another thing that can go stale.

Illustrative case study

A fictional logistics agent schedules dock appointments from a live calendar feed. After an outage, the feed silently serves fifteen-minute-old data, and the agent double-books two carriers into one slot.

The fix attaches a freshness timestamp to the state: if the calendar view is older than sixty seconds at decision time, the agent must refresh or queue the booking. The outage case becomes a defined path instead of an invisible assumption.

The invented lesson: model-based agents need to know the age of their knowledge. Stale state is a hazard with a timestamp.

Common misconceptions and mistakes

The model is a machine-learning model. In this term it means the agent's internal picture of the world; learned components may help build it but are not the point.
More memory is always better. State should hold what decisions need; bloated state is harder to keep fresh and correct.
Current input makes state unnecessary. Any decision depending on history, pending actions or hidden factors needs the accumulated picture.

Business use

Model-based design matters wherever the right action depends on history: conversations, scheduling, case handling and operations monitoring.

Specify three things for any such agent: what its state contains, how each field is refreshed, and what it must do when its information is stale or contradictory.

Check your understanding

A booking agent double-books a slot after its calendar feed goes stale. What state rule would prevent this?

Answer: A freshness rule: if the calendar view is older than a set limit at decision time, refresh or queue the booking rather than acting on stale state.

Related terms

Learn next

AI Agents and Automation for Small Business covers state, error alerts and safe retries in working agents: https://aiforbusinessandlife.com/academy/ai-agents-automation-small-business. [C4]

Sources

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

Previous: Model assertion Next: Model-based clustering