Every four months or so we try getting hyped on agents for ~ a week. Up to this point the answer has always been a pretty solid "nope this sucks back to manual code." The agentically coded parts of the codebase that came from these trials are palpably more annoying than the rest. However, the current trial is going better than the last ones have (Fable 5.1 really is astonishing). On the other hand, each other trial started with a honeymoon so it's up in the air.

It is actually pretty nice that now we can spin up an isolated copy of our whole system with one command line call- this was mandatory for letting agents run wild trying things, but mostly now it makes manual development and onboarding easier.