Ask a model to fix a failing test and you get a patch to paste. Ask an agent and it runs the tests, reads the failure, edits the file and runs them again until they pass or it gives up.
It can be the same model in both cases. The agent is the loop around it, plus the tools and
A safer prompt is not a sandbox.
A good agent sandbox is enforced by the host, not by vibes and wishful thinking. A model can be wrong and it can be tricked, it also can trick you.
Useful bar:
- files only where you grant them
- network only to allowlisted destinations
- a
The agents finished porting Campfire to Laravel and Django too. I'm sure there are optimizations that could be found there too, but the whole point is to get a sense of how frontier agents do with different environments out the box. github.com/basecamp/once-…
