What is Agentic AI? A Practical Introduction

Ask a chatbot a question and you get an answer. Ask an agentic system to get something done and it goes away, makes a plan, uses software tools, checks what happened, and keeps going until the job is finished or it needs your help. That shift, from answering to acting, is what people mean by agentic AI.

From chatbots to agents

A plain LLM call is one-shot: text goes in, text comes out. The model cannot look anything up, run anything, or change anything outside the conversation, and it forgets everything that is not in the prompt. That is perfect for drafting an email, and hopeless for “find out why our nightly build fails and fix it”.

An agentic system wraps the same model in a loop and gives it three things: tools it can call, state it can remember, and a goal it keeps working towards. For the build-failure task, an agent can read the logs, open the failing test, edit the code, re-run the tests, notice that one still fails, and try again. The model is the same; the system around it is what makes it agentic.

The agent loop

1. Plan. Given the goal and everything seen so far, the model chooses the next step. Some systems plan the whole task up front; many plan one step at a time.
2. Act. The model asks for a tool to be run: search the web, query a database, execute code, call an API, edit a file. The surrounding program runs it. The model itself never touches the outside world directly.
3. Observe. The tool result is appended to the model’s context, so the next decision can use it.
4. Decide. If the goal is met, answer. If not, plan again. A good loop also stops for other reasons: a step or cost budget, an error it cannot recover from, or a point where a human should decide.

What makes a system agentic

There is no sharp line, but systems that people call agentic usually show most of these properties:

• Goal-directed behaviour. It works towards an outcome, not a single reply.
• Tool use. It can take actions and fetch information. Our article on tool use and the Model Context Protocol explains how that works.
• Memory and state. Short-term memory is the context window; long-term memory is a store such as files, a database or a knowledge graph that survives between sessions.
• Planning and reasoning. It breaks a task into steps and revises the plan when something unexpected happens.
• Feedback and self-correction. It checks its own work, ideally against something objective such as a test suite, a validator or a simulation.
• Autonomy. It decides the next step itself instead of following a fixed script.

How much autonomy?

Autonomy is a dial, not a switch. A useful way to think about it:

1. Assistant. The model suggests and you act. Nothing runs without you.
2. Supervised agent. The agent acts, but asks for approval at checkpoints such as before it deletes, spends or sends anything.
3. Autonomous within guardrails. The agent works alone inside a sandbox with limited permissions and reports back.
4. Fully autonomous, long-running. It pursues open-ended goals for hours or days. This is rare, and risky.

Most systems that deliver value today sit at levels 2 and 3, and most mistakes come from giving an agent more autonomy than its reliability justifies.

Where agentic AI works well

• Tasks with cheap, objective checks. Code that has tests, data extraction with a schema, calculations that can be re-verified, simulations with numerical sanity checks.
• Multi-step work across several tools. Gather data from one system, transform it in another, write the result to a third.
• Exploratory tasks. Research, debugging and analysis, where the next step depends on what the previous step revealed.

Where it struggles

• Errors compound. If every step is 95% reliable, a ten-step task succeeds only about 60% of the time (0.95¹⁰ ≈ 0.60). Reliability per step matters far more than it does for a chatbot.
• Cost and latency. Every loop iteration is another model call, and the context grows as the agent works.
• Ambiguous goals. If success is hard to define, the agent cannot tell when it is done, and neither can you.
• Security. An agent that reads untrusted text, such as web pages, emails or documents, can be manipulated by instructions hidden in it. See how to evaluate and secure AI agents.

The vocabulary, untangled

• Generative AI is the broad family of models that produce text, images, code or other content. (Our generative models article covers the classic families.)
• An AI agent is a concrete system: a model plus tools, memory and a control loop.
• Agentic AI is the approach, and the property: building systems that act with some autonomy towards goals. You will also see agentic workflows, which are a more constrained cousin, covered in AI agents vs. agentic workflows.

A first experiment

The quickest way to understand agents is to build a tiny one. Pick a task that has an automatic check, such as making a failing test pass. Give the model three tools: read a file, run a command, and edit a file. Cap the number of steps, log every model call and every tool result, and require your approval for any edit at first. Watch where it succeeds and where it wanders; that is where your real design work is.

Further reading

• Yao et al., ReAct: Synergizing Reasoning and Acting in Language Models (2022), the paper behind the “think, act, observe” pattern.
• Anthropic, Building effective agents, a practical guide that distinguishes workflows from agents.
• On this site: agentic AI for science, with examples from my own research.