What is an AI agent? It's the chatbot that clicks for you
Ask a chatbot to find you a dentist near Tampines and it gives you a list. Ask an AI agent the same thing and it might open the clinic’s booking page, pick a Saturday slot and fill in your name. Same request. Very different amount of trust.
That gap is what the word “agent” is trying to name, and it is now stuck on so many products that it has started to mean nothing. So here is what it actually means, and what to ask before you let one loose.
A chatbot talks. An agent does.
Singapore’s Ministry of Digital Development and Information put it in one line when it launched its agentic AI framework in January: “Unlike traditional and generative AI, AI agents can reason and take actions to complete tasks on the behalf of users.”
The important word is actions. A chatbot produces text, and then you decide what to do with it. You copy the email, you click the link, you press send. An agent is connected to tools, such as a browser, your inbox, a calendar or a company database, and it presses the buttons itself.
Anthropic, the company behind Claude, draws a finer line in its engineering guide Building effective agents (December 2024). It separates workflows, where an AI follows “predefined code paths” a programmer wrote in advance, from agents, which “dynamically direct their own processes and tool usage”. A workflow is a recipe. An agent is a cook who decides the next step by tasting the pot.
The loop is the whole trick
Strip the marketing away and an agent is surprisingly plain. Anthropic’s own description: “They are typically just LLMs using tools based on environmental feedback in a loop.”
In practice that loop has four beats:
- Look. Read the task and whatever is on screen or in the file.
- Act. Use a tool: search, click, type, run some code.
- Check. Look at what actually happened. Anthropic calls this getting “ground truth” from the environment at each step, such as the result of a tool call.
- Repeat or stop. Decide the next move, or finish. Builders often add a hard limit, such as “a maximum number of iterations”, so it cannot wander forever.
The model underneath is the same kind of language model that powers an ordinary chatbot, still predicting text one token at a time. What changes is that some of that text is now an instruction to a tool, and the tool’s answer goes back into the next prediction. Intelligence did not suddenly appear. Hands did.
[[photo:1]]
The analogy: a runner with your card
Think of an agent as someone you send to the shops with your list and your bank card.
A chatbot is the friend who tells you which shop has the best price. Nothing leaves your wallet. The runner is more useful, but now three new questions matter. How much may they spend? When should they text you before buying? And what happens if a stranger at the market says “the owner told me you should buy this too”?
Every real risk with agents maps onto one of those three.
Spending too much, or on the wrong thing. The ministry’s announcement names “unauthorised or erroneous actions” as the new risk, because agents may be able to make changes “such as updating a customer database or making a payment”. Anthropic adds that autonomy brings “the potential for compounding errors”: a wrong turn at step two gets built on at steps three, four and five.
Never texting home. The ministry also flags “automation bias”, which it defines as “the tendency to over-trust an automated system that has performed reliably in the past.” The runner was fine the last ten times, so you stop checking the receipt.
The stranger at the market. This one has a name: prompt injection. Anthropic describes it as “malicious actors hide instructions in websites, emails, or documents to trick AIs into harmful actions”. In its own testing of Claude in Chrome, a planted email claimed emails needed deleting “for security reasons”, and before new defences were added, Claude “followed these instructions to delete the user’s emails without confirmation.” Across 123 deliberate attack tests, the attack success rate was 23.6% without the new safety measures and 11.2% with them. Better. Not zero. And that was the company’s own red-teaming, with the agent deliberately targeted.
[[photo:2]]
What Singapore wrote down
Singapore’s Model AI Governance Framework for Agentic AI, developed by IMDA and announced by Minister Josephine Teo at the World Economic Forum in Davos on 22 January 2026, is written for organisations deploying agents, not for households. Its four dimensions still read like good house rules for anyone:
- Bound it upfront. Pick suitable tasks and place “limits on agents’ powers such as agents’ autonomy and access to tools and data”. In runner terms, a card limit and a short list of shops.
- Make a human sign off. Define “significant checkpoints at which human approval is required.” Text me before you buy.
- Control it technically. Test it, and allow access only to approved services.
- Make sure users understand it. Transparency and training, so the person giving the instructions knows what the thing can do.
Under all four sits one sentence the ministry chose to stress: “humans are ultimately accountable.” If the runner buys the wrong thing on your card, the bank will not be calling the runner.
Products are starting to build in the same ideas. Anthropic says Claude in Chrome “asks users before taking high-risk actions like publishing, purchasing, or sharing personal data”, lets users grant or revoke access site by site, and blocks some high-risk categories of sites, including financial services. (A disclosure: Mentus AI’s own mentors run on Anthropic’s models, which is part of why we read its documentation closely. The same questions apply to any agent from any company.)
Three questions before you hand over the card
Next time a product says “agent”, skip the brochure and ask:
- What can it touch? Inbox, files, bank, calendar? Grant the smallest set that does the job.
- Where does it stop and ask? Anything you cannot undo, such as paying, sending or deleting, should need your tap.
- What will it read that a stranger wrote? Web pages and incoming emails are where planted instructions live. The more of those an agent reads unsupervised, the more a confirmation step matters.
And remember the advice in Anthropic’s own guide, which is unusually modest for a company selling this: find “the simplest solution possible”, which “might mean not building agentic systems at all.” Sometimes the friend with good advice is enough. Either way, you set the card limit before they leave, not after.
Was this useful?
Sources
- Singapore Launches New Model AI Governance Framework for Agentic AI · Ministry of Digital Development and Information, Singapore
- Building effective agents · Anthropic
- Piloting Claude in Chrome · Anthropic