How is this different from a greedy agent?

A greedy agent takes the best immediate step; a problem-solving agent searches sequences of steps toward the goal.

Does it need machine learning?

No. Classic problem-solving agents search with a given world model; learning may improve the model or the heuristics.

When does planning fail?

When the world model is wrong, the goal is vague, or the environment changes faster than plans complete.

AI Glossary / Agents and automation / Intermediate

Problem-solving agent

Definition

A problem-solving agent is an agent that works toward a goal by searching: it imagines sequences of actions, evaluates where they lead, and chooses a sequence that reaches the goal. It plans before it acts.

What it means

A reflex agent asks: what do I do now? A problem-solving agent asks: what sequence of steps gets me to the goal? The difference is foresight. The agent considers futures, compares them and commits to a plan.

The classic picture needs four things: a goal, a model of what actions are available, a way to predict what each action does, and a measure of progress. With those, the agent can search the space of possible action sequences for one that works.

Search is the honest core of the idea. The agent does not know the answer; it explores: try this step, see where it leads, backtrack, try another. Cleverness lies in exploring promising paths first.

The approach shines when the world is predictable and the goal is clear: puzzles, routing, scheduling. It struggles when the world is unpredictable or the goal is fuzzy, which is why real systems mix planning with reaction and human judgment.

For a business reader, the problem-solving pattern explains both the power and the cost of planning features in AI tools: exploring futures takes computation, and the plan is only as good as the model of the world it was made with.

How it works

The agent formulates the problem: the current state, the goal test, the available actions and the cost of each. Then it searches a tree of possibilities, where each branch is an action and each level is a step into the future.

Search strategies differ in how they explore. Uninformed strategies explore systematically without knowledge of the goal's direction. Informed, or heuristic, strategies use an estimate of distance to the goal to try the most promising branches first.

The output is a plan: a sequence of actions predicted to reach the goal. The agent executes it, usually watching for divergence between prediction and reality, replanning when the world refuses to match.

Limits are practical: the search space explodes with problem size, so real systems bound depth, prune dominated options or accept a good-enough plan rather than the perfect one.

Three practical examples

Routing: compare sequences of turns against the map and the clock before setting off, rather than turning at every corner by instinct.
Scheduling: explore orderings of jobs to find one that meets all deadlines, instead of booking each job greedily.
Configuration: search combinations of components to find one satisfying all compatibility rules within budget.

Think of it this way

Think of planning a journey with a map spread on the table. You trace a few routes with a finger, compare distance, tolls and traffic, and only then start the car. The tracing is the search; the driving is the execution.

The limit: the map can be wrong, and roads close. A plan made on a bad map fails confidently, which is why execution monitoring and replanning are part of the design, not extras.

Worked example

planning a delivery sequence

Suppose a courier has fifteen parcels and a hard promise to finish by 6 p.m. A greedy approach visits the nearest address each time and risks finishing late with a long tail across town.

A problem-solving approach formulates the full problem: locations, time windows, driving times, the deadline. It searches orderings, using the map to estimate each sequence's finish time, and picks one that beats the deadline with margin.

Midday, a road closure invalidates part of the plan. The agent re-searches from the current state with the closure added, rather than pushing through the old sequence. The scenario is invented; formulate, search, monitor, replan is the cycle.

The measure of success is the day's outcome, met deadline and total miles, not the elegance of any single choice.

Illustrative case study

A fictional events company uses a planning agent to schedule its venue crews. The plans look optimal on paper and fail in practice because the model of the world omits setup overruns, which happen every weekend.

Rather than tuning the search, the team fixes the model: historical overrun margins become part of each task's expected duration. Plan quality jumps because the futures being compared are now honest.

The invented lesson: a problem-solving agent is hostage to its world model. When plans fail, inspect what the planner believes about the world before blaming the search.

Common misconceptions and mistakes

Planning means knowing the future. Planning compares predicted futures; predictions from a wrong model produce confident bad plans.
Search always finds the best answer. Real searches are bounded and accept good-enough solutions when the space is too large to exhaust.
More planning is always better. In fast-changing environments, long plans rot before they finish; short horizons with frequent replanning often win.

Business use

Problem-solving approaches suit stable, well-modeled decisions: routing, scheduling, configuration, resource allocation and what-if analysis.

Give any planning tool three things: a current and honest model of your constraints, a measure that reflects real cost, and a habit of replanning when reality deviates from the plan's assumptions.

Check your understanding

A crew-scheduling planner produces optimal-looking plans that fail every weekend. What should be inspected first?

Answer: The planner's world model: whether task durations and constraints reflect reality, including weekend overruns, before tuning the search itself.

Related terms

Learn next

AI for Business Owners covers choosing AI approaches that match the stability of your problem: 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: Problem generator Next: Procedural attachment