Hacker News
new
top
best
ask
show
job
Code as Agent Harness
(
arxiv.org
)
5 points
by
matt_d
8 hours ago
1 comment
promptsaredead
2 hours ago
I think a lot of SOTA models are already going this way. Long autonomous tasks in claude code / codex already do this to stay on track and avoid multiplying errors over many steps.