Observability is not enough.

Gartner predicts that 40% of organizations deploying AI will adopt dedicated AI observability tools by 2028 [1]. Having visibility into how your AI performs is worth it, no question, but the more interesting line in that same report is what's actually pushing the shift: "there's a growing need for predictive issue detection and real-time actionable insights in AI models." Teams are buying observability, but it seems that what they need is prediction and action.

Observability tools were really built for a person to sit down and evaluate runs and data, but like a lot of AI use cases, that person tends to become the bottleneck and there's so much more we could be handing off to automation.

After hearing from many AI engineers about how they work with observability tools, it sound like the current stack assumes someone will do the legwork - running the evals, comparing the runs, reading through the charts and scores. In practice, teams hook their agents up to a platform for visibility, but an engineer still has to go in, figure out what went wrong, and ship the fix themselves. This means that maintenance is reactive most of the time. For example, quality drifts until a customer complains, latency slips until someone catches it, and costs creep up until finance starts asking questions. The tooling could see all of it, but nobody was watching for those specific problems.

It doesn't have to work this way anymore. The same capability shift that created agentic workflows applies to keeping them healthy them. An agent can cluster failures across thousands of runs, track down the root cause, weigh the quality-latency-cost tradeoffs, and even draft the code change to implement a fix. Teams deployed agents to take humans out of the loop of their business workflows, but a human is still in the loop when it comes to fixes.

What still can't be automated away — and shouldn't be.

Nobody wants an agent silently rewriting production prompts. The right model is one a CTO described to me without me even asking. He called it Depend-A-Bot. The bot combs the system continuously, finds something worth fixing, and shows up with the change and the evidence. A human approves it. The judgement part stays with a human, but all the work leading to that decision gets automated. The human time spent shifts from interpreting a chart someone to a reviewable change someone only has to sign off on.

The new way to maintain agents.

Observability answered "what is my agent doing?" and laid the answer out well for human eyes. The next layer answers "what should I change?" and doesn't need you to read a chart to get there. You can keep the dashboard, but know that it doesn’t do the work for you.