What has actually held up after a year of agentic AI in DevOps?

Deploy summaries and pull request review look like practical places to start. Summaries of merged changes can make releases easier to follow, while review tools may catch missing configuration changes or deployment risks that are easy to overlook.

Neither replaces a reviewer. Large pull requests can generate a lot of noisy comments, so tuning what the tool reports matters too.

I'd be cautious about natural language to Terraform. A convincing demo isn't the same as a correct infrastructure change, and reading every generated line can eat up any time saved drafting it.

There is a verification angle worth exploring, though. Terraform plans and policy assertions could check some properties of a proposed change. That doesn't prove the whole change is safe or eliminate review, but it might make particular checks repeatable.

Alert correlation seems useful as a triage aid, not as an incident commander. A summary can cut through a pile of related alerts and suggest a probable cause, but the suggestion can be wrong.

That distinction matters for automatic rollbacks. An anomaly isn't necessarily a broken deployment, so using the same uncertain signal to take action is a much bigger step than showing it to an engineer.

The useful boundary for now is authority. Let an agent read, draft, or summarize, then have a person check anything consequential. A chatbot that confidently suggests a bad production command can be dangerous even if it never runs the command itself.

I'd keep production changes behind human approval rather than assuming a polished answer is a trustworthy one.

Don't forget to count the tool itself as another service to operate. A pipeline optimizer has to save enough time to justify its setup and babysitting, not just produce an impressive recommendation.

For lower stakes work, troubleshooting assistance, ticket drafts, documentation, and infrastructure diagrams are other possibilities to evaluate. Existing infrastructure may be harder to fit into agent workflows than an environment designed around them.