Agents
Agents are finally becoming good enough for day-to-day use.
With good models they're already quite productive with minimal nudging. Without cost subsidization through subscriptions they get insanely expensive though.
Where they're useful
The sweet spot, for me, is well-scoped work with a tight feedback loop — the kind of task where you can describe the outcome precisely and verify it cheaply. Give an agent a clear spec and a way to check its own work, and it will grind through the boring parts without complaint.
Give them tools, not just text An agent is only as capable as what it can reach. Most of the leverage comes from Agent Skills — small, composable capabilities — rather than from a cleverer prompt.
The catch
The bill arrives quietly, then all at once.
Run an agent in a loop and the token cost compounds fast. A subscription hides that; pay per call and you feel every retry. So the discipline is the same as with any expensive tool: keep the scope narrow, and stop when it's done.