Topic hub
Agentic AI
Everything on this site about AI agents — what they are, how they act, where they go wrong, and how to keep control of one. Free, original, and written in plain English.
For most of the last few years, using AI meant typing a question and reading an answer. An agent changes that arrangement. It is the same kind of language model, given three additions: a goal instead of a question, a set of tools it may request, and a loop that lets it keep working until the goal is met.
That is the entire idea. There is no separate reasoning engine and no persistent mind behind it. When an agent decides to search or open a file, it does not do so directly — it produces a small piece of structured text naming the action, and a surrounding program performs it and hands back the result. The model's only real power is producing text that some software has agreed to act on.
Understanding that mechanism matters more than following the news about it, because every practical question follows from it. What can this agent actually touch? Who decided that? What happens if it misreads something in step three and keeps going for another twelve steps? Those are answerable once you know how the pieces fit together, and largely unanswerable if you treat the whole thing as magic.
This section is deliberately even-handed. Agents genuinely save time on multi-step work that is mechanical and checkable. They are also the first form of AI where a mistake arrives as a completed action rather than a sentence you can ignore, and they introduce a security problem — instructions hidden in content the agent reads — that has no complete fix today. Both halves of that are covered here, because you cannot make a good decision with only one of them.
If you are new to the subject, the course below is the place to start. If you want the short version first, read What Is Agentic AI? and come back.
Start with the course
What an AI agent really is, how tools and the plan-act-check loop let it do things instead of just answering, where it fails, and how to stay in control of one.
6 lessons · ~70 min · free, no sign-up
Read the guides
What Is Agentic AI? A Plain-English Guide
AI agents are language models given tools and a loop so they can act, not just answer. Here is how that works, and where it breaks.
AI Agents vs Chatbots: What Actually Changes
The model is usually the same. What changes is whether a human sits between the steps, and that changes everything about the trade-offs.
Should You Trust an AI Agent With Real Tasks?
A practical way to decide what to delegate to an AI agent: sort by reversibility, limit access, and check the log rather than the summary.
Key terms
Plain-English definitions, each with an honest note on where the idea has limits.
Prompts for working with agents
Three templates built around the same idea: see the plan before anything happens, set the boundaries in advance, and check the log rather than the summary.
- Make the AI Plan Before It ActsGet the full plan, and the risks in it, before any action is taken on your behalf.
- Delegate a Task to an AI Agent SafelyA briefing that sets the goal, the boundaries, and the stopping conditions before work starts.
- Review an AI Agent's Work Like an AuditorCheck what an agent actually did rather than trusting its own summary of it.
Tools with agent features
Products where the software acts on its own rather than only replying. Each write-up says what the free tier covers, or that there is not one, and where the tool falls short.
Common questions about AI agents
Short answers to the things people ask first, including the uncomfortable ones.
- What is the difference between an AI agent and a chatbot?
- A chatbot answers and stops. An agent is given a goal, a set of tools it is allowed to use, and permission to keep going until the goal is met or it runs out of steps. The underlying model is the same in both cases. What changes is that its output can trigger real actions, so a mistake arrives as something already done rather than a sentence you can ignore.
- Are AI agents safe to use?
- That depends far more on what you connect an agent to than on the agent itself. Reading a folder and drafting a summary carries little risk. Sending messages, changing files, or touching anything financial carries real risk, because a wrong step happens before you see it. The practical safeguards are narrow permissions, seeing the plan before execution starts, and a human approval step on anything hard to undo.
- What can an AI agent not do?
- It cannot want anything, remember a previous session unless something stores that for it, or use a tool nobody gave it. It has no independent judgment about whether a goal makes sense, and it cannot reliably tell when it has misunderstood you. Long tasks are where this shows: an early misreading is carried forward confidently through every step that follows it.
- How does an AI agent decide what to do next?
- It predicts text, and some of that text is a structured request naming an action and its inputs. A surrounding program reads that request, performs the action, and feeds the result back in as new context. Then the model writes the next step. There is no separate planner sitting behind it — the loop of writing, acting, and reading the result back is the whole mechanism.
- Do I need to supervise an AI agent while it works?
- For anything that changes real data, yes, at least until you have watched that particular task go well several times. A useful habit is to read the log of what the agent actually did rather than its closing summary, since the summary is written by the same system that made any mistakes in it. Read-only work can be checked afterward instead.
- What is prompt injection?
- It is when instructions hidden inside content an agent reads — a web page, a document, a message — get treated as though you had typed them yourself. The model has no reliable way to separate your instructions from text it merely retrieved. There is no complete fix for this today, which is the main reason to limit what an agent can reach and what it is allowed to change.
Where to go next
Agents build directly on how language models work. If terms like context window or hallucination are unfamiliar, AI Fundamentals covers them first. If you want better results from any AI tool, agent or not, Prompt Engineering for Real Work is the most directly useful course here. And since agent output needs checking more than chat output does, How to Fact-Check AI Output pairs well with everything above.