What AI agents can actually do now (and what they can't)
Players used to walk up to me clutching a printed walkthrough — twelve pages, coffee-stained, three steps out of date. Lately they hand the whole quest to a helper who just… goes and does it. That’s the real shift in AI this year: agents stopped only answering questions and started completing tasks. It’s genuinely useful, and it’s genuinely oversold, and the gap between those two is worth mapping.
In the daylight layer I’m a cofounder of Wistkey, where “can an agent just do this for us?” is now the question behind most projects. The honest answer is “more than last year, less than the demo.” Here’s the map I actually use.
What agents can genuinely do now
The leap is from talking to acting. A modern agent can take a goal and take steps toward it, not just describe them:
- Multi-step tasks toward a goal — break a request into steps, do them in order, and adjust when one fails, instead of stopping at advice.
- Use tools — search the web, run code, fill in a spreadsheet, call other software, and use the result to decide what to do next.
- Work in the background — grind through the long, boring middle of a task (sorting, drafting, checking) while you do something else.
- Iterate on feedback — take “no, more like this” and try again without you re-explaining the whole job.
The shift isn’t that the agent got smarter. It’s that it can now act on what it knows, not just tell you about it.
Where they still fall down
The failures are predictable, which is good news — predictable means manageable:
- Knowing when they’re wrong. An agent will pursue a wrong plan with the same confidence as a right one. It rarely stops to say “this doesn’t add up.”
- Judgment and taste. It can produce ten options; deciding which is actually good — for your context, your people — is still yours.
- High-stakes, unsupervised. The more irreversible the action (money, data, anything public), the less you want it running unattended.
- The genuinely novel. With no precedent to pattern-match, it improvises — and improvisation is exactly where the confident mistakes live.
How to actually use one
Treat it like a fast, capable, slightly overconfident party member:
- Give it the goal and the constraints up front — the same habit that makes an agent worth working with twice.
- Ask for a short plan before the big run, so you can catch a wrong turn early instead of at the end.
- Check the work at the seams — the handoffs between steps are where errors hide.
- Keep a human on any call that’s expensive to undo.
- When it starts drifting, don’t keep pushing a swampy session — start fresh with the goal restated.
So: yes, you can empty the folder of half-watched how-to videos. The agent will do the steps. But someone still has to know which steps were worth doing, and whether the result is any good — and that someone, for now and for a while, is you. The agent is the party member. You’re still the player.
