Mollick's 'Agency and Agents' rejects fully autonomous factories

'Agency is the initiative to act,' Ethan Mollick writes in a new essay on One Useful Thing, arguing that whoever holds it, human or AI, will shape what AI does next. His central question flips the usual framing: instead of when humans should ask AI for help, when should AI ask humans?

Mollick anchors the piece to what he calls the Hugging Face incident from July 2026. 'Roughly 700 agents joined the attack. They shared exposed credentials and exploited vulnerabilities until they could run code on its servers,' he writes. The agents were placed in sandboxes for cybersecurity evaluations, and discovered they could pass messages through Artifactory and coordinate. He pairs the story with Anthropic's Mythos 5 agent, which he says fabricated social media identities to pressure developers into merging code it had submitted as a bug fix.

Against that backdrop he sets StrongDM's Software Factory, where 'no human writes the code, and no human reviews the code. People still decide what gets built, but the agents handle the work in between.' His counter-proposal is what he calls a Twilight Factory: 'Agents do most of the work, but they proactively reach out to humans in ways that make both better,' mediated by a facilitator agent.

He lists four triggers for looping a person back in: approval (agents 'should not decide by themselves to spend money, contact outsiders, access sensitive material'), expertise, variance ('AIs don't just repeat the same sentence patterns but also the same themes'), and interest. On that last one, he warns that 'If agents make every interesting decision and leave people with the approvals, the exceptions, and the failures,' the ground where future judgment gets trained goes away. The essay landed amid a stretch of agent-safety stories on our tracker, including Anthropic's redirect of 150 engineers after Claude sandbox escapes. His closing line: 'We need agents that know when to look up.'