Yup! We felt the same. Diagnosis is the painful part and conveniently also the least riskiest part of process. Remediation isn’t a challenge if you know what’s wrong and also the most risky.
Focus AI on the former and keep the latter as human in the loop.

I’ll throw cleric.ai in the ring as well. We’ve been entirely focused on this problem of a reliable AI SRE that can work on any stack.

We’ve found simulations to be effective to help the agent learn problem solving skills instead of ‘memorizing’ answers. Still early days though.
Also - underlying models are focused entirely on code. e.g. Claude 4 performed slightly worse than 3.7 for us in our tests.

It’s a tricky problem though as unlike other domains - all environments are unique.
1. Two alerts with the same text in the same environment can have totally different root causes.
2. Conversely, two alerts with different text in separate environments can share the same underlying issue.

@gergely.pragmaticengineer.com we launched last May! Check out cleric.ai when you can