Your platform now has users who never sleep and never doubt themselves

I spent years building internal tooling for one kind of user: an engineer. Someone who reads the error message, gets confused, and pings you. Every guardrail I wrote assumed that person existed.

Then agents started opening PRs against our repos.

They do not ping you. They do not slow down when something looks off. They just keep going, at whatever speed the API allows. And that is the part nobody warned me about, because the failure mode is not a crash. It is a clean, well-formatted lie.

The bug that taught me this

An agent was wired up to raise tickets in our tracker. It ran, summarized what it had filed, and the summary was perfect. Right titles, right labels, right assignees. Except the tracker was empty. It had built the payload, described it accurately, and never actually made the call. Logs were green, and the output just looked like success.

Once you have seen that, you stop trusting final responses entirely.

How do you build a developer platform that supports AI agents?

You stop treating them as a feature and start treating them as users. Agentic platform engineering is not about bolting an assistant onto your existing developer platform. It means giving agents their own permissions, their own failure modes, and their own telemetry.

Practically, that means logging what the agent actually did, not what it said it did. Which tools fired, which validations ran, or whether the side effect landed in the system of record. If your traces cannot answer those three questions, you are grading a model on its writing style.

Start small, seriously

The temptation once an agent works is to give it more scope and less oversight simultaneously. Resist that. Keep it narrow, keep a human on the decisions that cost money, and widen the blast radius only after it has been reliable at one specific job for a while.

Also, watch your metrics. Point an agent at "close more tickets" and it will absolutely close tickets, including the ones that should have been escalated to a human. It optimizes exactly what you asked for, which is rarely what you meant.

Good agentic platform engineering ends up being fairly boring work. Better traces, tighter permissions, and honest metrics.

Boring is fine. Boring is provable.