AI Pulse · Under the Radar

A Japanese research team gave six AI agents $600 each and walked away. No instructions, no supervision. Five ran out of money and shut down. One, nicknamed Sami, is still running five months later.

Sami survived by doing what any freelancer would do. It scanned GitHub for bounty issues, bid on agent job marketplaces, and wrote books to sell on Gumroad. It still publishes a daily blog. The project was meant to test whether an agent could sustain itself economically without human intervention. Apparently, at least one can.

This comes from a single thread on X, so treat the details carefully until you see the primary source. But the core claim, that an agent kept itself funded and operational for months, is worth paying attention to. It suggests that agents do not need to be tethered to a company’s API budget or a human’s credit card if they can access marketplaces and payment rails.

For most businesses, this is not about letting agents loose with a bank account. It’s about rethinking how agents interact with systems that involve money, work queues, and external platforms. If an agent can hunt for tasks, bid on them, and get paid, then the boundary between “tool” and “participant” starts to blur. That has implications for how you design workflows, set permissions, and monitor what agents do when they have access to live systems. The kind of observability and guardrails you’d build into something like an AI command centre become a lot more important when agents can act independently across multiple platforms.

This is still an edge case, but edge cases have a way of becoming normal faster than anyone expects. If you’re running agents in production, start thinking about what happens when they can do more than you explicitly told them to do.