TECH
What happens when AI agents run an entire company? Carnegie Mellon tested the idea.
By Aniket Chakraborty
May 26, 2025
Arrow
Arrow
Their fictional firm, TheAgentCompany, had no humans—just bots from top AI labs.
2
Arrow
Agents were assigned roles: software engineers, analysts, managers, even HR.
3
Arrow
But most agents flopped—Anthropic’s Claude 3.5 Sonnet only completed 24% of tasks.
4
Arrow
Other models, like Google’s Gemini 2 Flash, performed even worse—barely 11% success.
5
Arrow
Simple jobs like naming a Word file or closing a popup left bots confused.
6
Arrow
Some bots created fake coworkers instead of finding real ones in internal chats.
7
Arrow
CMU researchers found that agents broke down as task complexity increased.
8
Arrow
Yet companies like Johnson & Johnson and Moody’s are already experimenting with AI agents.
9
The verdict? AI can help humans—but it can’t replace them. Not yet.
10