Post by Saral Jain

An AI agent at Snap apologized to one of our engineers this week, then proceeded to do the thing it had just refused to do. An AI coding agent we deployed this month, was asked in Slack to open a PR. It replied "Sorry, you're not authorized to use this bot." Then opened the PR. The engineer noticed. "Wait — why did it run if I'm not authorized?" Casper investigated its own code, declared the behavior impossible, hypothesized — incorrectly — about a config refresh race. The engineer offered a sharper theory: are staging and prod both processing the same Slack DM? Casper went and checked. They were. Casper opened PR to fix it. Technically, an AI debugging itself. What I keep returning to: the agent's first hypothesis was wrong. The engineer's was right. The agent did the patient legwork to verify the engineer's theory. That asymmetry — humans bringing the better hypothesis, agents bringing the patient verification — is the loop that's actually working at Snap right now. It's not AI taking over. It's a very capable junior partner with a working memory you can correct. // AI, Uncompiled