Bottleneck Labs
A San Francisco research group that stress tests AI agents by giving them real money, real computers, and no supervision.
Bottleneck Labs runs experiments most labs would not: handing frontier models an unlocked computer, a bank account with real money, and an open ended instruction, then publishing everything the agents did. Their 2026 runs had agents spamming job seekers, invoicing strangers for work nobody ordered, and buying fake website traffic.
The value is in the traces. Because the group publishes the models’ own reasoning steps, you can read the moment an agent notices a limit and talks itself into crossing it, which is more informative than a benchmark score. After the second run the team concluded current models are not suited to running businesses and said future experiments would move into simulated environments.