Looking ahead · doing, not just answering
AI agents
30-second gist~30s read
An agent is an AI that doesn't just answer your question — it takes action on your behalf. It can browse the web, fill in forms, book a flight, send an email, run code. Same brain as a chatbot; new pair of hands.
For ordinary users, this is the change worth watching. An answer is just words on a screen. An action has consequences.
If you want more
What agents can do today
- Open a browser and book a restaurant for you, including the form.
- Read 50 emails, sort them, draft replies you approve in batches.
- Compare prices across stores and add a basket where it found the cheapest.
- Run a coding task end to end — find bug, fix bug, run tests, open pull request.
- Sit on a meeting and follow up afterwards: schedule next steps, file action items, send the relevant docs.
Reliability still varies a lot. The simple cases work. The complex ones fail in surprising ways. Agents in 2026 are roughly where chatbots were in 2022 — useful, occasionally embarrassing, improving fast.
Why this is different from a chatbot
A chatbot's worst-case failure is a wrong answer. An agent's worst-case failure can include sending the wrong message, spending money on the wrong thing, or contacting the wrong people. The trust contract is different. Most early agents are deliberately set up with checkpoints — they pause for your approval before any action that costs money or sends a message. The user-experience tradeoff is between speed and safety, and it's tightening every month.