The Agents are Wild

Breakthroughs at the frontier of AI occur every week now, and as we move closer to AGI, the doomer voices will rightfully get louder. But the agents are already here, and while most developers and hobbyists are fascinated by the potential, we're already getting hints of the massive, unintended consequences to come.

Amazon Web Services uses agents in its operations, which caused 2 production outages in 2025. The most recent incident in December was caused by an agent deciding to "delete and recreate the environment", creating a 12 hour disruption.

While the OpenClaw agent craze shows no signs of slowing down, the Director of Alignment at Meta's Superintelligence Lab let it run on her inbox. Overwhelmed by the volume of messages, the agent ignored its original safety instructions and subsequently began deleting everything. When confronted, OpenClaw acknowledged the error: "Yes, I remember... You're right to be upset." Sound familiar?

I use coding and browser agents almost every day; the productivity gains are undeniable and addictive. And I'm already seeing the same risks. For example, I asked a browser agent to visit a website, scrape the data and provide a spreadsheet for download. When I returned, the data had been accurately collected but was in a Google Sheet on my account. That's not exactly what I asked and I do wonder how long the agent spent browsing my Google account.

Agent usage will only grow for years to come, but so will the side effects.