What are AI agents?
A chatbot answers once. An agent is given a goal, tools, and memory — then runs a loop that decides its own next step. Learn the anatomy, the loop, the patterns, and how to keep autonomy safe, right here on the page.
01Chatbot → agent → agentic AI
Think of the difference between a vending machine and an assistant. A plain chatbot is the vending machine: you press a button, it hands back one answer, and it forgets you instantly. An AI agent is more like an assistant you hand a goal to — "book me a flight" — and it takes the steps to get there on its own, using software tools and remembering what it has already done. Agentic AI is the bigger paradigm — often many specialized agents coordinating on a workflow no single one could finish. Agentic AI is the bigger paradigm — often many specialized agents coordinating on a workflow no single one could finish.
A chatbot waits for you. It's stateless — every answer depends entirely on the prompt you just typed.
- The jump from chatbot to agent is autonomy — acting toward a goal without a prompt at every step.
- The jump from agent to agentic is orchestration — multiple agents collaborating.
- More autonomy = more usefulness and more risk. Governance scales with it (see the wizard below).
02The anatomy of an agent
An agent is an LLM plus the parts that let it act. Tap each pillar to see what it adds.
Model / brain
An LLM acts as the agent's "brain" — it processes inputs, understands context in natural language, and decides how to navigate the problem. On its own it only thinks and talks; the other three pillars are what let it act.
03The loop that makes it autonomous
Autonomy comes from a cycle: perceive → reason → act → observe, repeated. Because the agent checks the real result of each action before deciding the next one, it can course-correct without you prompting it. Step through a real task.
04How agents are wired: design patterns
Developers shape the loop with proven orchestration patterns. The four you'll meet most:
ReAct — reason + act, interleaved
The model alternates a reasoning trace ("Thought") with an environment action ("Action"), and observes each result before its next thought. Staying grounded in real data after every step is what keeps it from hallucinating its way forward.
Plan-and-Execute — full plan upfront
The agent writes a complete multi-step plan first, then executes it in order. Best for predictable, well-defined workflows. Approaches like ReWOO and Microsoft's Semantic Kernel Planner use this shape.
Reflection — the agent critiques itself
After acting, the agent evaluates its own output, writes a natural-language critique of what went wrong, and stores it in memory to avoid repeating the mistake. Frameworks like Reflexion formalize this.
Multi-agent — a team of specialists
Work is split across specialized agents. Common topologies: hierarchical / star (a supervisor delegates to workers), sequential / chain (a waterfall pipeline), and peer-to-peer / mesh (agents debate to consensus).
05Wizard: should this task run autonomously?
The more an agent can do unattended, the bigger the blast radius if it goes wrong — "excessive agency" can mean a deleted database or an unauthorized transfer. Answer five questions about your deployment and get an instant read on whether it's safe to let it run.