Agent capabilities for legal work
Legal work is where nearly right is wrong.
harvey-labs is a benchmark built to evaluate and improve agent capabilities for supporting legal work. The reason it exists is that a plausible answer and a defensible one look identical until someone checks.
01
Evaluate
Give the same task to different agents and get a number back, so a claim about capability has something behind it.
02
Improve
The project puts improvement in its own description, which means the score is meant to feed work rather than end an argument.
03
Supporting, stated plainly
The description says supporting legal work. That word is doing real scoping, and a design should keep it rather than round it up.