What is an AI agent?

An AI model put to work in a loop: it decides, takes an action, looks at the result, and repeats until the job is done. A chatbot talks and stops. An agent acts until the work is finished.

Yesterday I handed one a messy export of last month's orders and asked it to find the refunds we issued by mistake. It pulled the file, cross-checked each refund against the order, found four that shouldn't have gone out, and wrote me the list with the reasons. I didn't tell it the steps. It figured out the steps.

That loop is the whole idea. The thing that made it useful wasn't that it was smart. It was that it kept going.

What it is (in plain English)

Start with the model. A large language model on its own predicts text and then stops. You ask, it answers, it sits there. That's a chatbot. Useful, but it has no hands and no follow-through. If you're weighing the steered chat window against a system that runs on its own, see AI agents vs AI assistants for which one fits which job.

An agent puts that same model in a loop with tools. It decides what to do, takes an action (search a file, hit an API, send an email), looks at what came back, and decides the next step from there. It keeps cycling until the job is done or it hits a stopping point.

The cleanest way to hold this in your head: an AI agent is a model plus a harness. The model is the brain. The harness is the scaffolding around it that gives it the tools, the loop, the memory, and the guardrails. This isn't my pet theory; it's the standard frame practitioners use (Hugging Face literally writes it as "Agent = Model + Harness"). The reason two products built on the same model can feel like completely different tools is the harness, not the brain.

It also explains the difference between an agent and the automation you already run. If a human had to anticipate and hard-code every step, it's automation. If the model decides the steps at runtime based on what it sees, it's an agent. Your Zapier flow or RPA bot follows a script you wrote. You own the plumbing. With an agent, the model owns the plumbing. It picks the path while the work is happening, which is exactly why it handles the messy refund file that would break a fixed script the moment a row looked different.

Why CEOs care

The chat window is the slow version of AI. You're the copy-paste layer between the model and the result. An agent removes you from that loop, so work happens while you're in a meeting or asleep.

The bigger reason is the model is becoming a commodity and the harness is where the leverage lives. Pretty much every CEO has a paid Claude or ChatGPT account now. Almost none of them have an agent doing real work against their own files and systems. The gap between those two CEOs compounds every week, and it's invisible from the outside until the lead is enormous. The skill worth building isn't picking the smartest model. It's learning to put one to work in a loop on a job that actually matters in your week.

Where you'll see it

• A coding agent like Claude Code, the clearest live example: it reads your codebase, edits files, runs the tests, reads the errors, and fixes them. The whole observe-act-repeat loop in front of you. See What is a coding agent?.
• A customer-service agent that reads a ticket, looks up the order, checks your refund policy, issues the refund, and replies, choosing different tools for different tickets.
• A lead-followup agent that researches an inbound lead, drafts and sends a personalized reply, books the meeting, and logs it to your CRM.
• A back-office agent that reconciles invoices against bank transactions and flags the exceptions a fixed RPA script would choke on.
• A research agent that runs searches, reads, follows the thread, and writes you a sourced brief, deciding on its own when it has enough.

Most of these reach you through a product: Claude Code, ChatGPT's agent mode, Salesforce Agentforce, Microsoft Copilot agents. Underneath, they're all a model in a loop with tools. When you wire one to your own systems, those tools usually come through MCP.

What to do next

Be honest about the limits before you trust one too far. Agents compound their own errors: a small mistake early in the loop snowballs through every step after it. They also have a finite context window, so long multi-step jobs overflow and the agent starts to forget what it was doing. That's why agents look incredible in a two-minute demo and fall apart on a real eight-hour task, and why every reliable one in production today runs narrow, with a human checkpoint before anything irreversible.

So start narrow. Pick one repetitive, well-scoped job you already understand cold (reconciling something, triaging an inbox, pulling a weekly number) and watch a coding agent run it end to end one time. Don't automate it yet. Just watch the loop work. That single observation will change how you think about your whole week. Tell me what it did.