AWS principal engineer says the best engineers are ‘paranoid and lazy’ when it comes to AI

David Yanacek, a senior principal engineer at AWS, believes the most effective engineers are both "paranoid" and "lazy". For Yanacek, paranoia involves being highly vigilant about potential safety risks, while laziness drives him to seek ways to prevent these risks and automate processes. Despite the common belief that judgment and taste are key traits of top engineers in the AI era, Yanacek thinks paranoia and laziness are equally important.

According to Yanacek, who works on agentic AI and safety, paranoia ensures he remains alert to possible dangers when granting an AI agent permission to perform tasks without oversight. Laziness, on the other hand, means he wants to avoid manually executing repetitive tasks.

In the current workflow, AI has largely taken over coding tasks. However, engineers used to spend significant time waiting for manual tests, code approvals, and design tweaks. Now, AI agents handle most of this work, with human intervention reserved for design discussions and complex reviews.

These loops involve generating code, running tests, addressing issues, and updating the system for deployment. While one loop can take hours, Yanacek can work on other projects or add features to products in the meantime.

For efficient and time-saving agentic workflows, agents need to operate autonomously. Yanacek spends much of his time determining which tools agents can use without supervision, which actions require approval, and where to enforce these limits. To achieve this, he ensures his "auto-trust" file, which specifies which tools and commands an agent can use independently, is well-defined.

Yanacek's primary focus is on AI safety, and this mindset allows him to sleep well at night. He believes that engineers need a combination of paranoia and laziness to be alert to potential issues when agents act autonomously while also building reliable systems that minimize the need for constant supervision. The ultimate goal is to establish the agent's trustworthiness, enabling it to automate tedious tasks without human oversight.