Learn › AI Agents: Foundations › Types of AI Agents

Types of AI Agents

Beginner 🕐 11 min Lesson 6 of 7
What you'll learn
  • Describe the five main agent architectural types and their tradeoffs
  • Identify which architecture suits a given task
  • Explain why most real-world systems are hybrid agents

A Taxonomy of Agent Architectures

Agents differ not just in what they do, but in how they're built. The architectural type you choose determines how the agent reasons, what it can learn, how it scales, and where it'll break. Here are the five main types you'll encounter.

Reactive Agents

Reactive agents respond to their current perception without maintaining a model of the world or any memory of past states. They map inputs directly to outputs using rules or patterns.

Roboticist Rodney Brooks pioneered this architecture in the 1980s with his "subsumption architecture" — robots that behaved intelligently by layering simple reactive rules, not by building an internal world model.

  • Strengths — fast, predictable, easy to debug, good for latency-sensitive tasks
  • Weaknesses — cannot reason about the future; makes suboptimal decisions when context from the past matters
  • Best for — real-time systems, simple automation, tasks where the right response can always be determined from the current state alone

Deliberative Agents

Deliberative agents build an explicit model of the world and use reasoning or search to plan sequences of actions toward a goal. They think before they act.

  • Strengths — high-quality decisions, can reason about consequences, handle complex multi-step planning
  • Weaknesses — computationally expensive; slower to respond; can struggle in fast-changing environments where the world model goes stale
  • Best for — tasks where planning quality matters more than response speed (strategy, research, complex problem-solving)

Learning Agents

Learning agents improve their behavior over time based on feedback. Rather than relying on predefined rules, they adjust what they do based on outcomes — through reinforcement learning, fine-tuning, or in-context learning from examples.

  • Strengths — improve with experience; can handle tasks that are hard to specify with rules
  • Weaknesses — require careful oversight; without proper controls they can learn unintended or biased behaviors
  • Best for — tasks where the right behavior is easier to demonstrate than to specify, and where feedback signals are clear

Multi-Agent Systems

Multi-agent systems consist of multiple agents working in a shared environment. The agents may cooperate (each handling a different subtask), compete (optimizing against each other), or operate independently on the same problem in parallel.

  • Strengths — parallelism, specialization, robustness (one agent failing doesn't stop the whole system)
  • Weaknesses — complex to design and debug; emergent and sometimes unpredictable behaviors; require coordination protocols
  • Best for — large tasks that can be decomposed, tasks requiring specialized sub-agents (a researcher + a writer + a reviewer), or scenarios where redundancy matters

Hybrid Agents

Most real-world agents combine architectural types. A common pattern is a hybrid that pairs reactive low-level control (handle immediate interrupts quickly) with deliberative high-level planning (reason about the goal and sequence of steps). Autonomous vehicles are the classic example: they react to obstacles in real time while also planning a route.

In LLM-based agents, hybrid architectures often look like a fast reactive layer that handles common patterns with rules, plus a slower deliberative layer that invokes the full LLM only when needed. This reduces cost and latency without sacrificing quality on complex tasks.

Key takeaways
  • Reactive agents respond to current state only — fast and predictable, but no memory or planning
  • Deliberative agents build a world model and plan — high quality decisions but slower and computationally heavier
  • Learning agents improve with feedback — powerful but require careful oversight to avoid learning unintended behaviors
  • Multi-agent systems distribute work across specialized agents — parallel and robust, but harder to design and debug
  • Most production systems are hybrid: a reactive layer for speed plus a deliberative layer for quality on hard tasks
Learn more every week Get 5 hand-picked prompts + new AI tools every Monday. Free.