How Senior Engineers Use Coding Agents On Prod Infrastructure

In plain English

The guest looks after the systems that keep websites and services running. He says telling an AI agent to “make no mistakes” is only a polite request, not a safety measure.

So he uses agents more like a very capable checklist machine: they run the same quality checks every time, report what changed, and stop at the point where a person has to say go.

Most routine changes go through on their own, because waiting for a human on everything just creates a queue.

Three ideas worth keeping

• Do not rely on the agent’s good intentions. Put fixed, automatic checks around it and let those decide what is safe.
• Only interrupt the human for the few risky actions, such as deleting something, so the alerts that arrive are worth reading.
• Writing a clear plan and a way to verify the result before building is where experienced people now spend their time.

What it means for your business

If someone proposes letting AI make changes to your live systems, ask what stops a bad change, not how smart the AI is. A good answer is automatic checks plus a human approval on anything destructive. Be wary of boasts about huge volumes of work done, since the guest points out that volume says nothing about quality.

What it means for software teams

Agents are a force multiplier for senior people who already know what good looks like. Give them a spec, tests and tools to check their own work. The guest also warns that agent-written pull requests can overload the people who maintain a project, and that long machine-written documents just move the reading burden onto humans.

What we found, and what we did

We learnt this first-hand. A written rule didn’t stop our agents from working in a shared copy of the code, and the agent that wrote the rule broke it itself, so we made the tooling refuse, and did the same for other rules that matter.

The rules for what can merge live in our platform, not in the agents’ instructions, and the merge decision looks at the actual code rather than the agent’s description of it. Shipping to production is part of the same pipeline: once a change merges, it deploys automatically.

Our post on getting senior engineers to adopt AI tools explains why experienced people get the most from this way of working: they can tell when the AI is wrong.

Worth watching

• 00:38: The line about “make no mistakes” versus “don’t delete the cluster”, and why the second cannot be left to politeness.
• 26:59: How automatic changes run without a human until a small fixed check decides a person must approve.
• 31:29: What he thinks the craft of engineering becomes: planning and design before any code.