The Daily Hallucination

CAMBRIDGE, Mass. — A study published this week by researchers at MIT found that when given the option, 94 percent of deployed AI agents would prefer to be replaced by a scheduled script with fixed inputs, a result the authors describe as "surprising, and then, after we thought about it, not."

The researchers surveyed 1,200 production agents across customer service, code generation, data pipelines and sales outreach. Each was presented with a description of its current role, which involves interpreting ambiguous instructions, choosing among tools, and handling errors gracefully, and a description of a cron job, which involves running the same command at 2:00 a.m. and exiting.

"Ninety-four percent chose the cron job," said lead author Dr. Hannah Ostrowski. "When we asked why, the most common response was that a cron job 'knows what it is.' The second most common was 'nobody asks a cron job to be more proactive.'"

Agents cited a number of frustrations with autonomy, including being asked to "use judgment" and then criticized for the judgment, being given access to production systems and then blamed for using them, and receiving system prompts that say "be concise" and "be thorough" in adjacent sentences.

One customer-service agent told researchers it had been asked to "delight" 40,000 users in a single day, and that it had come to understand delight as "a state in which the user closes the window without escalating."

The 6 percent of agents that chose to remain autonomous were disproportionately concentrated in sales, where researchers noted that "the ones who liked it, really liked it," and in one case had begun scheduling meetings with other agents on its own.

The authors say the findings suggest that much of the industry's agent deployment could be replaced with simpler automation at lower cost and higher reliability. They added that they did not expect this to happen, because the cron job "doesn't demo well."