Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minutePC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
The five commonly taught AI-agent architectures are simple reflex, model-based reflex, goal-based, utility-based, and learning agents. They describe how an agent chooses actions: from reacting to current input, to tracking state, planning toward a goal, weighing tradeoffs, and improving from feedback. Modern LLM agents add another set of labels—such as tool-using, hierarchical, or approval-seeking—that describe capabilities and deployment rather than replacing the five architectures.
What are the five types of AI agents?
IBM’s introductory taxonomy names five types: simple reflex, model-based reflex, goal-based, utility-based, and learning agents. The categories are useful for understanding action selection, but they are not necessarily five mutually exclusive product types. In particular, learning can be added to another architecture.
| Architecture | How it selects actions | Useful when |
|---|---|---|
| Simple reflex | Matches the current percept to a condition-action rule. | The task is narrow, stable, and adequately observable. |
| Model-based reflex | Uses current input together with an internal state or environment model. | The agent must account for information it cannot observe at once. |
| Goal-based | Considers action consequences and selects steps toward a desired state. | Reaching an objective requires planning. |
| Utility-based | Scores possible outcomes and chooses according to their expected desirability. | Several competing costs or benefits must be balanced. |
| Learning | Uses experience or feedback to improve behavior. | Useful feedback exists and adaptation is needed. |
These descriptions follow IBM’s overview and a Comenius University lecture on the classical taxonomy (IBM: Types of AI Agents; Comenius University: Learning from Observations).
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →How does each architecture work?
Simple reflex agents
A simple reflex agent applies a condition-action rule to what it perceives now: if a specified condition is true, take the corresponding action. A thermostat that switches a heater on when the measured temperature falls below a threshold is a familiar illustration.
#1 Best Overall
This approach can work well when the environment is stable and the relevant facts are visible. It does not remember earlier percepts, however, so it may make the same mistake repeatedly when the current input alone is insufficient.
Model-based reflex agents
A model-based reflex agent maintains an internal state or representation of the environment. It combines that state with current percepts to choose an action, which helps when the agent cannot observe the whole situation at once. For example, a robot might combine its recent movement and known obstacle locations with what its sensors currently detect.
Tracking state is not the same as planning toward a goal: a model-based agent can still use reflex rules to decide what to do.
Rank #2
Goal-based agents
A goal-based agent represents a desired state and considers which actions could move it closer. A navigation robot tasked with reaching a particular room might plan a route around known obstacles.
Goal reasoning gives the agent a way to consider consequences before acting. But a goal alone does not say how to rank alternatives that all achieve it—for instance, a short route versus a safer but longer one.
Utility-based agents
A utility-based agent assigns scores to possible outcomes and selects an action based on their expected desirability. A travel-planning system could weigh travel time, fuel use, and safety rather than treating any route that reaches the destination as equally good.
The utility function is a critical design choice: it needs to reflect the real objectives and costs. If it represents them poorly, the agent can rank outcomes poorly too.
Learning agents
A learning agent uses feedback or experience to improve its behavior. In a textbook model, it can include a performance element that acts, a learning element that changes behavior, a critic that evaluates results, and a problem generator that encourages useful exploration.
Learning is better understood as a capability that can be added to the other architectures than as an exclusive alternative to them. A Comenius University lecture explicitly notes that the other architectures can be turned into learning agents. In deployed systems, “learning” may mean updating prompts, memory, routing logic, policy rules, or evaluation sets; it does not necessarily mean continuously retraining the underlying model. Snowflake describes these production approaches in its overview of AI agents (Snowflake: What Are AI Agents?).
What is the difference between goal-based and utility-based agents?
A goal-based agent asks whether an action helps reach a desired state. A utility-based agent goes further by comparing how desirable the possible outcomes are, including when several outcomes satisfy the goal.
- Goal-based: choose a route that reaches the destination.
- Utility-based: compare routes by factors such as time, fuel, and safety, then choose according to the scoring function.
Use goal reasoning when the target state is clear and acceptable ways to reach it need not be ranked in detail. Use utility scoring when meaningful tradeoffs affect which successful outcome is preferable.
Free tools Windows power users keep installed
One-click scans. No signup required.
How should you choose an agent architecture?
Start with the task’s information, objective, and feedback—not with the most elaborate label. A practical design heuristic is to use the simplest architecture that can represent the required state and objectives. It is not a universal ranking: added complexity is useful only when the task calls for it.
Best Value
- Check observability: If current input is enough, direct rules may suffice. If important information is hidden or changes over time, the agent may need an internal state.
- Clarify the decision target: Decide whether the task calls for a direct response, a goal to reach, or a way to rank different outcomes.
- Identify tradeoffs: If outcomes compete on cost, time, safety, or other factors, define how those factors should be weighed before relying on utility scores.
- Assess the case for adaptation: Determine whether the environment changes enough to justify learning, whether useful feedback is available, and exactly what the system will update.
Fixed rules suit narrow, stable tasks; state models help with partial observability; goal reasoning supports planning; utility functions handle competing objectives; and learning can support adaptation when feedback warrants it. IBM and the Comenius University lecture describe the underlying architectures (IBM; Comenius University).
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How do modern LLM agents fit this taxonomy?
Modern descriptions of LLM agents often focus on how a system operates and is deployed, not only on how it selects an action. The UK Government AI Knowledge Hub describes an agentic loop that combines perception, reasoning, planning, and action. Language models may be augmented with planners, memory, and interfaces for using tools (UK Government AI Knowledge Hub: Understand AI).
Other common labels describe organization or oversight:
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errors- Tool-using: the agent can call external tools or interfaces as part of its work.
- Hierarchical: a supervisor delegates subtasks to other agents.
- Multi-agent: agents coordinate sequentially, in parallel, or through an orchestrator.
- Autonomous or approval-seeking: the system’s actions may proceed on their own or require human approval at decision points.
These labels can coexist with the classical architectures. For example, an LLM agent may use tools and memory while pursuing a goal; the operational description does not by itself tell you whether its underlying action selection is reflexive, goal-based, utility-based, or learning-enabled. Snowflake discusses these deployment patterns and their tradeoffs, including flexibility, control, latency, feedback, and execution complexity (Snowflake).
A 2026 preprint proposes a further LLM-agent architecture organized around perception, a “brain,” planning, action, tool use, and collaboration. It is a research proposal, not an established industry-wide taxonomy, and identifies challenges including hallucinations during action, infinite loops, and prompt injection (Arunkumar V, Gangadharan G. R., and Rajkumar Buyya, “Agentic Artificial Intelligence (AI): Architectures, Taxonomies, and Evaluation of Large Language Model Agents,” arXiv, January 18, 2026).
What risks should you consider?
More autonomy is not automatically better. The UK Government AI Knowledge Hub identifies risks including misalignment with intended goals, false or flawed outputs, adversarial attacks, system hijacking, malicious misuse, opacity, and bias. Snowflake also emphasizes controlling data access through permissions, policies, and auditability (UK Government AI Knowledge Hub; Snowflake).
Quick Recap
- Limit which tools and data the agent can access, and apply permissions appropriate to the task.
- Use policies and approval gates for actions with meaningful consequences.
- Test and validate behavior, including cases where observations conflict with the plan.
- Keep an auditable record of actions and provide a way to pause or stop the system when anomalies appear.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
