# What AI agents actually do (and what they still can’t)

> A chatbot answers when you talk to it. An agent takes the steps itself. What agents do well today, where they still fail, and how to use one this week.

By The WYDM Team · 18 Oct 2026 · Topic: [The Agentic Web](https://wydm.app/blog/topics/the-agentic-web) · Tags: [AI agents](https://wydm.app/blog/tags/ai-agents)

Canonical: https://wydm.app/blog/what-ai-agents-actually-do

# What AI agents actually do (and what they still can't)

"AI agent" has replaced "AI" as the word everyone uses and almost nobody defines. A chatbot answers when you talk to it. An agent is different in one load-bearing way: you give it a goal, and it takes the steps itself. It searches, it compares, it uses tools, it comes back with the thing done or a good reason it could not be.

That difference sounds small and changes everything about what the technology is for. Here is my version of both halves.

## What an agent does

An agent holds a goal across multiple steps without you steering each one. "Find me a weekend flight under £200" is an agent task: it searches, filters, compares times, and returns three options with prices. "Write me a poem about my dog" is a chatbot task, one step, one answer, no world to act on.

The good agents today do narrow, verifiable work well. Booking, scheduling, searching, monitoring, drafting from material you already have, summarising a pile of documents, chasing a status. The pattern in the successes is the same: clear goal, checkable result, limited room for creative interpretation.

## What an agent still can't

![ai agent vs chatbot - a helper walking toward a goal](https://media.wydm.app/blog/0ea8dcaaf82b0e2e8ca40b660fc98c7e.webp)


Anything where "done well" is a matter of taste, and anything where a wrong action costs real money or a real relationship. An agent can draft the email; the sending should stay yours. [Stanford's Human-Centered AI Institute describes agents as systems that plan and use tools towards goals](https://hai.stanford.edu/ai-definitions/what-is-agentic-ai), and I think the human-centred part of that framing is the half worth remembering. The systems are capable long before they are reliable, and that gap is where the trouble happens.

Trust an agent with reversible steps and small stakes. Watch it closely on anything that spends, sends or publishes. The people getting value out of agents treat them like a competent intern: real work, real supervision, nothing irreversible without a sign-off.

## The dating example, because it is a good one

Dating is where agents get interesting, because the grunt work is enormous and the stakes are personal. The searching, the profile reading, the small talk before the interesting talk, the scheduling Tetris. A lot of it is admin sitting between two people who might like each other.

That is the design bet behind [WYDM's agent system](https://wydm.app/): agents that handle the admin of dating, searching within your circle, vetting, opening conversations, getting two people to the same cafe on Saturday, and a profile that stays genuinely yours. The agent never gets the last word on anything. Meeting, deciding, and turning up stay human jobs, and the published agent API is scoped exactly to those admin steps.

## How to use one this week

Pick a task you can check in ten seconds, give an agent three tries at it, and grade the results yourself. You will probably land where I did. Agents are already worth using for the boring middle of a job and still worth supervising for the parts that matter.