AWS Engineer: Best Developers Are 'Paranoid and Lazy' on AI

By Olivia Hart

The best engineers working with AI agents today are "paranoid and lazy," according to David Yanacek, a senior principal engineer at AWS who works on agentic AI and safety. He made the claim in an interview with Business Insider, and he meant it as a compliment.

The pairing sounds like a punchline. Yanacek treats it as a working method. Paranoia keeps an engineer alert to what an unsupervised agent might do; laziness supplies the motivation to make sure it never comes to that.

What does 'paranoid' actually mean here?

"Paranoid means I'm very sensitive to what could possibly happen if I gave an agent permission to do a thing without supervision," Yanacek told Business Insider. He is not describing anxiety. He is describing a habit of enumerating failure modes before granting an agent any measure of independence.

Laziness is the other half. Yanacek said he does not want to perform repetitive tasks manually, and that reluctance pushes him to build systems that handle the repetition instead.

The skeptical read: this is a neat reframing of old engineering virtues — caution plus automation — with fresher vocabulary. What Yanacek does not say is how many failures his approach has actually caught, or how often his agents misbehave despite it. The claim rests on his own practice, not on incident data.

How has the workflow changed?

Yanacek said AI has taken over most coding tasks, but he argued that coding was never the real bottleneck. Before agents, engineers spent days waiting on manual tests, code approvals, and design tweaks — waiting that depended on teammates' reviews.

Now agents run much of that work in loops:

• generating code,
• running tests,
• reworking problems,
• preparing updates that get merged and deployed.

A single loop can run for hours. During that time, Yanacek can shift to other projects or add features to a product. Human intervention, he said, is now reserved for design discussions and complex reviews.

The caveat is obvious and he acknowledges it implicitly: hours of unsupervised machine activity only works if the guardrails were set correctly beforehand. That burden now sits with the engineer.

What are the guardrails?

For agentic workflows to save time, agents need autonomy, Yanacek said. So he spends much of his time deciding which tools agents can use without supervision, which actions require approval, and where those limits are enforced.

The instrument at the center of this is what he calls the "auto-trust" file: the set of rules specifying which tools and commands an agent can use independently, and which need a human sign-off.

"I always think about what is in the auto-trust file, because that's how you get productivity: by having the agent be able to do things without you having to micromanage every single command that it runs," Yanacek said.

His stated goal is a fleet of agents making the right calls without constant review and approval. That requires clear policies for what an agent should and shouldn't do — a policy-writing exercise, not a coding one. Notably, the senior engineer's core job has drifted from producing software to writing rules about software behavior.

Does paranoia scale?

"Most of what I think about is AI safety," Yanacek said. "And because I'm doing that, I sleep pretty well."

Engineers, in his framing, need a layer of paranoia and laziness so they stay alert to what could go wrong when an agent acts without supervision — while simultaneously building systems reliable enough that supervision of every action is unnecessary.

The end goal, he said, is to do enough work up front to make an agent trustworthy enough to run unsupervised, so it can automate the tedious parts.

In plain words: Yanacek is betting that judgment, expressed as policy, is what lets autonomy scale. Whether "auto-trust" files hold up as agents grow more capable is the question his sleep depends on — and the one this interview cannot answer.