Edition 2026
The best AI agents
An agent is only interesting if it finishes the job without a human re-reading every step. That is a much higher bar than a chat window, and most products marketed as agents do not clear it.
So this list cares about what happens on the second hour of a long task: whether it recovers from a failed tool call, whether it stops when it is unsure, and whether you can see what it did afterwards.
What we look at
- Completes a multi-step task end to end without step-by-step approval
- Recovers from a failed tool call instead of looping or inventing
- Leaves an auditable trace of every action it took
- Stops and asks rather than guessing on irreversible actions
- Published, per-task cost you can predict before you run it
Helps if you have it: A recorded run of a real multi-step task, with the trace and the final cost visible.
Nothing in this list yet. If you know something that belongs here, add it. Name and a link is enough.
Add an entry