What Is an AI Agent, and How Is It Different from a Chatbot?

In this article＋

01 / ARTIFICIAL INTELLIGENCE

The boundary is an action loop, not a tone of voice

An AI agent is not a chatbot with better manners. It is a goal-driven system with tools and a feedback loop; the real question is who can see a wrong step and stop it.

OpenAI's practical guide treats this loop as central. A long prompt without observation and revision may still be a chatbot behind ambitious branding.

You have probably seen the demo: an agent opens ten tabs, does a little dance and returns a result. It looks great. My first question is less cinematic—where did it quietly go wrong?

A system that only reads flight data is not in the same risk class as one that can edit a CRM and send a refund email. The interface may differ by one button; the consequences do not.

02 / ARTIFICIAL INTELLIGENCE

Tools let text change the world

Reading a calendar retrieves information; sending email, issuing a refund or changing a database creates consequences. Once tools are available, an error can expand from a wrong sentence into correctly executed wrong actions.

Agent capability is reasoning multiplied by the authority and systems attached to it.

For a Hong Kong business, the painful part is not a one-minute delay. It is the wrong email, record or refund already sent before anyone notices.

A common Hong Kong mistake is asking, 'Can it be fully automatic?' I prefer the less glamorous question: which step is safe to delegate, and which step must stay human? That is usually where the budget survives.

03 / ARTIFICIAL INTELLIGENCE

The overlooked capability is stopping

Demos celebrate an agent completing ten steps and rarely show when it should stop. Dependable systems need maximum steps, cost limits, repetition detection, low-confidence exits, permission boundaries and approval.

Without an exit condition, autonomy may mean only that a system can travel farther from the goal before anyone notices.

I start with read access, then drafts, then carefully gated writes. Less theatre. More reversibility.

Demos love to show an agent completing ten steps, then quietly skip the moment when the data is stale at step three. A dependable design needs step limits, cost ceilings, loop detection, low-confidence exits and explicit permission boundaries.

04 / ARTIFICIAL INTELLIGENCE

Memory carries one mistake into the next step

Agents preserve state, tool results and intermediate beliefs. If a tentative inference is stored as confirmed fact, later actions build on a false foundation.

A good trace is more than a debug log. It maps which choice the model proposed, which tool executed it and which person approved it.

05 / ARTIFICIAL INTELLIGENCE

Not every workflow needs autonomy

Fixed, well-specified and low-exception processes are usually cheaper and easier as ordinary automation. Agents help when steps cannot all be listed beforehand, work requires judgement over unstructured input, and tool feedback can correct the next decision.

Autonomy added only for an AI appearance is expensive decoration.

06 / ARTIFICIAL INTELLIGENCE

A simple test

Ask four questions: does it have a goal, select its next step, use tools, and change plan from results? The more answers are yes, the more agentic it is.

Then ask the decisive fifth: when one step goes wrong, who can see it and stop the run?

Autonomy can be adjusted one permission at a time

Autonomy is adjustable, not binary. An agent may read freely but require approval before writing, plan independently but confirm external actions, or automate only below a value limit.

Permission, step count, sensitivity and reversibility can each create a gate. The goal is autonomy where errors remain observable and recoverable.

Human approval can become decorative

Human approval can become ceremonial. If a confirmation says only allow tool without showing recipient, amount, changes and evidence, people learn to approve reflexively.

Effective human-in-the-loop design presents the facts most likely to be wrong at the decision point and permits correction, not only accept or cancel.

Agency moves the evaluation target

A chatbot can be judged mostly by its response. An agent must be judged by the trajectory: tool choice, permissions, intermediate state, stopping and final side effects.

A run that ends correctly after unsafe or wasteful steps is not necessarily successful. This is the hidden shift introduced by agency.

Evaluation moves from “was the sentence good?” to “was the entire path authorised, efficient, observable and recoverable?

”

07 / ARTIFICIAL INTELLIGENCE

Agent difficulty behaves like a multiplier, not an addition: model judgement multiplied by tool permissions, multiplied again by the number of steps it may take.

A system that only reads flight data is not in the same risk class as one that can edit a CRM and send a refund email. The interface may differ by one button; the consequences do not.

08 / ARTIFICIAL INTELLIGENCE

The least glamorous feature is the stop button

Without an exit condition, autonomy merely means the system can drift away from the goal for longer.

09 / ARTIFICIAL INTELLIGENCE

Questions people actually ask

Does any tool use make a system an agent?
No. Rules plus an API can be automation. An agent chooses its next step from the result instead of following a completely fixed script.

Should it be allowed to write to systems?
Start with read access. Gate writes, payments, deletion and outbound messages one by one, with an audit trail.

Where should a team start?
Choose a low-risk, reversible task with a clear goal but messy inputs, such as triaging enquiries or comparing documents.

Sources

Sources support the mechanisms and limitations discussed here. Models, products and prices change; check the official page and date when a detail matters.