What Is an AI Agent? Loop, Tools & Examples

The short definition (steal this)

Friday refresh note for this pillar: we rewrote the examples around weekday duties founders actually run — price checks, briefings, rank notes — and we killed the abstract taxonomy filler. If a paragraph did not help you decide buy vs skip, it is gone.

An AI agent is software that takes a goal, uses tools (browser, files, APIs, search), and runs a loop — think, act, observe, repeat — until the goal is done, blocked, or paused for a human.

A chatbot replies. An agent proceeds. If the system cannot leave the text box, call it chat. Marketing decks that rename autocomplete as “autonomous agents” are why founders waste a quarter testing the wrong thing.

One more distinction that saves budget: copilot vs agent. A copilot sits beside you in the editor and waits for the next keystroke. An agent leaves with a brief and comes back with artifacts. Both are useful. Confusing them is how teams buy a chat seat when they needed a weekday duty.

We are not precious about the word. We are precious about the loop. For the ops-angle version of this split, see AI agent vs chatbot for operations. For how a real cloud browser changes the ceiling, bridge to CloudAxis: giving AI agents a real cloud browser.

Agent vs chatbot vs RPA

Same Tuesday, three different machines.

System	Input → output	World access	Fails when…
Chatbot	Prompt → text	None (or toy tools)	You needed a click, not a paragraph
RPA bot	Trigger → scripted clicks	Known UIs only	The page layout shifts
AI agent	Goal → multi-step loop	Tools: browser, files, APIs	Bad grounding, tool loops, no human gate on irreversible acts

Zapier-style automation still wins when the API is stable and idempotent. Agents win when the UI is hostile, the path is fuzzy, or the output needs a short judgment call. We tested that split in Can AI replace Zapier? We tested it — hybrid is the boring, correct default.

The loop, with numbers

Goal Jess actually needed: “Every weekday by 8:00am, open these three pricing pages, note any change vs yesterday, drop a six-line brief in Slack.”

A chat model can describe the workflow in 40 seconds. An agent has to run it: navigate, extract, compare to yesterday’s file, write the brief, deliver. In our own competitor-monitor setups, a clean weekday run is usually 3–8 tool steps per URL — not one heroic prompt. Miss a subdomain on day one (Jess did; we have too) and the first brief is wrong. Fix the URL list once. By week two the morning message is boring. Boring is the product.

If you want that work while you sleep, schedules matter more than model brand. Start with why your AI agent should work while you sleep, then climb to CloudAxis for OS-level scheduling depth: how to schedule AI agents 24/7.

What changed by late 2026

Two years ago “agent” demos were live browser sessions with a human whispering the next click. Useful theater. Bad operations.

What actually moved for small teams:

Hosted cloud browsers — full Chrome, not text scrapers, so JS-heavy pricing pages stop being a dead end. Product lane on CloudyBot; mechanism depth on CloudAxis’s browser piece.
Scheduled duties — the agent comes back Monday without you opening a laptop. That is the difference between a toy and a worker.
Hard caps — metered APIs punish exploration. Caps pause at a ceiling. Live numbers live on pricing; the philosophy is in hard caps vs pay-per-use. Free is still the right on-ramp (no card). Paid plans bill immediately and stop at the limit.

What did not magically get solved: grounding drift, CAPTCHA walls, and humans pasting production passwords into group chats. Agents make those failures louder. That is a feature if you watch the run log.

Five checks before you buy an “agent”

Print this. Bring it to the vendor call. Or skip the call and answer it from docs in 20 minutes.

Can it pursue a multi-step goal without you prompting each step? If every action needs a new message, you bought chat.
What tools exist — and are they real? A cloud browser that executes JavaScript beats a “browse” badge that returns markdown summaries. See cloud browser AI automation and how CloudyBot works.
Where do files and memory live between runs? Yesterday’s pricing sheet must still exist Tuesday. Reset-every-session is a chatbot with amnesia.
What happens at the usage ceiling? Pause beats surprise overages. Compare ceilings on pricing.
Where is the human gate for irreversible acts? Sending email, changing live prices, moving money — those need approval, not vibes.

Contrarian stance you can disagree with: for a sub-20 person company, a hosted agent with hard caps beats a self-hosted gateway 9 days out of 10. The tenth day is when you have a platform engineer who enjoys 2am restarts. Fair guide: hosted AI assistant vs self-hosted agent. Product overview on the parent brand: CloudyBot on CloudAxis.

If your real job is competitor pages or rankings, skip the philosophy and go straight to competitor monitoring or SEO monitor — then wire the duty in the dashboard.

Where agents still fail (and why that is useful)

Wednesday afternoon. A duty hit a cookie banner that covered the price table, declared “no price found,” and wrote a cheerful empty brief. That is not sabotage. That is an honest miss. The useful part: the run log showed the screenshot. A human spent four minutes teaching a dismiss selector. Thursday’s brief was correct.

Expect these failure classes in 2026:

Tool loops — the model retries a dead click. Caps and step limits matter more than pep talks.
Weak grounding — it invents a number instead of reading the DOM. Demand artifacts: screenshots, CSV rows, quoted text.
Context bloat — pasting whole HTML dumps into the thread. Summarize; do not hoard.
Missing human gates — anything that spends money or publishes externally should wait for a yes.

Vendors that hide failures behind “it usually works” are selling demos. For the OS-level view of duties that keep running overnight, see CloudAxis on always-on agents. Vendors that show the browser, the file, and the pause-at-cap behavior are selling operations. If your evaluation week does not include one intentional failure, you are not evaluating — you are watching a trailer.

If you only remember one sentence: an AI agent is a goal-chasing loop with tools — and if your “agent” cannot produce a file, a browser trace, or a scheduled run without you babysitting every click, you bought a chatbot with better adjectives.

FAQ

Is ChatGPT an AI agent?

Mostly no. ChatGPT answers and can use limited tools in a session, but it is still chat-shaped: you start it, it stops when you stop. A production agent keeps a goal, uses tools across steps, and can run on a schedule without you staring at the tab. If you are comparing team chat seats to ops agents, start at our ChatGPT alternative page.

How is an AI agent different from RPA?

Classic RPA follows scripted clicks on known screens. Change the DOM and it breaks. An LLM agent reasons about what it sees and can replan when a page shifts. RPA wins on stable, high-volume internal UIs; agents win on messy public web and judgment calls.

What should I automate first with an AI agent?

Pick work that is repetitive, checkable, and reversible: competitor price snapshots, SEO rank notes, research digests. Avoid irreversible sends and payouts until you have a clean track record and a human gate.

Why do hosted AI agents use hard caps?

Multi-step browser work burns tokens and minutes unevenly. Hard caps pause the service at a ceiling so a stuck loop cannot become an invoice surprise. See CloudyBot pricing and hard caps vs pay-per-use.

Do I need to self-host to run a real agent?

Only if you want the ops tax. Hosted assistants give you a cloud browser, schedules, and caps without babysitting a gateway. Self-hosting can make sense for strict data residency or custom stacks — compare hosted vs self-hosted before you buy servers.

What Is an AI Agent? Definition, Loop & Examples (2026)