Will AI take my DevOps job? One year later

A year ago I wrote Will AI take my DevOps job?. My answer back then: parts of it, yes. The repetitive parts. But not the context, the responsibility, or the talking to people. I ended with "AI proposes, humans decide."

Looking back is more interesting than making predictions. Here's what actually happened.

What I predicted

In short, I said:

1. AI would write more of the boring configuration.
2. AI would get very good at reading logs.
3. Agents would start doing small fixes on their own.
4. Context, responsibility and communication would stay human.
5. A new part of the job would be building guardrails for AI agents.
6. MLOps would become more important.

What actually happened

Boring configuration: yes, completely. Writing YAML from scratch already feels old-fashioned. Reviewing it is the job now.

Reading logs: yes, and more. AI connected directly to observability tools, through MCP and vendor integrations, changed the first minutes of an incident. Instead of me searching, I ask questions and check the evidence.

Agents doing small fixes: faster than I expected. Agents that read a failing build, find the cause and open a pull request went from a demo to a normal part of the workflow for many teams. Context and responsibility: still human. I haven't seen any AI take ownership of a production decision, and I don't want to. The incidents I remember from this year were all decided by people. AI made the investigation faster, not the accountability lighter.

Guardrails: this became real work. Permissions for agents, sandboxes, review rules, deciding which tools an AI may call. This took more of my time than I expected, and it's very DevOps: building safety rails around powerful automation.

MLOps: yes. Everyone wants AI features, and someone has to deploy, monitor and pay for them. GPUs, model serving and LLM observability became normal infrastructure topics.

The job didn't shrink. It moved up a level, exactly as I hoped, but faster than I expected.

What I got wrong

• I underestimated how quickly agents would improve.
• I didn't expect security, especially prompt injection, to become such a big part of the AI conversation.
• I thought adoption would be slower in conservative companies. In some places it was, in others it was faster than anything I've seen.

The guardrails I built

The practical lessons from building guardrails this year:

• Agents work on branches, never directly on main.
• No production credentials for AI tools, ever.
• Narrow, short-lived tokens per task.
• Branch protection and required checks apply to bots like to people.
• Read-only access first, write access only where it's proven useful.
• Log every action an agent takes.

None of that is new. It's least privilege and good CI, applied to a new kind of user.

What I'd tell someone starting today, updated

The advice from last year still stands: learn the fundamentals, use AI daily, practise the human skills. I'd add two things:

• Learn to review well. Reading a diff critically and quickly is now one of the most valuable skills in DevOps.
• Learn a bit of security. Understanding permissions, supply chain risks and prompt injection makes you the person teams trust with AI tools.

So, did AI take my job?

Not yet. It took some of my tasks, and honestly, I don't miss them. What's left is more interesting: designing systems, making decisions, and keeping AI on a safe track.

My take, one year later: AI proposes, humans decide. The proposals got much better. The deciding is still mine.

— Desly