Hacker News
new
top
best
ask
show
job
I ran each AI coding-agent sandbox's own tests against the others
(
github.com
)
2 points
by
carlostapia
4 hours ago
2 comments
arvind_ravichan
an hour ago
this becomes a tiff of war between agents each with their own guardrails. how do you arbitrate in such cases? and where do you stop? there is no end on critiquing an agent.
carlostapia
4 hours ago
[flagged]