Jediah Katz argues agent harnesses still matter as models absorb coding patches

Jediah Katz, a harness engineer at Cursor and Grok Bot, argues that smarter models are absorbing the coding patches that agent harnesses once supplied, but that harnesses still matter.

According to a Chinese summary of his AI Engineer New York talk posted on X by meng shao, Katz says harnesses remain useful for opening new task types, for infrastructure, for trust and safety, and for efficiency.

Katz also predicts that within six months, 90% of agent work will run in the cloud rather than on personal computers. The summary is secondhand and quotes Katz's talk, not a transcript.