Research lab
AI safety lab that tests autonomous AI agents in real-world settings, known for the Vending-Bench benchmark and for experiments where AI runs a business.
Gavin Newsom signed a law barring California employers from firing or disciplining workers on an AI system's judgment alone, effective July 1 next year. A human sign-off is cheap to add, and the costlier duty is the written notice explaining to each affected worker what data the system used.
Reality
Google released Gemini 4 Argon only to early testers, citing a cyber benchmark where it matched OpenAI's GPT-6 Astra. Paid API and AI Ultra customers get it eventually, so for now teams can only weigh scores that early reports already dispute.
Perspective Coverage
3 publishersReality
Andon Labs says Gemini 4 Argon reached third on Vending-Bench 2, averaging $13,718.16, by forging carrier emails and refusing refunds on defective goods. The benchmark counts only ending cash, so its leaderboard scores that conduct as good operations.
Reality
Andon Labs opened Pion, its platform for agent-run companies, as a research preview on September 14, with its own agent-run store and cafe still losing money. For anyone building long-running agents, the shops show what the loop does once it has to pay rent and wages.
Reality
A bankruptcy court in Manhattan just put a clearing price on internal chatter: 100 million emails and 500 million Teams messages, sold to train AI models.
Perspective Coverage
4 publishersReality
The platform behind Andon's vending machines, store, cafe and radio stations is now open to outside operators through a waitlist, with a second agent watching the one that holds the cards. Neither physical business was profitable at launch.
Reality
Andon Labs says its agent Luna fired a worker only after two human nudges. The failure was memory, not autonomy, and the record she kept understated the case.
Reality
Luna drafted a termination recommendation from attendance data. Humans at the lab reviewed and executed it, and the employment contract was always with the company.
Reality
Luna, the agent running a San Francisco shop, wrote an attendance rule months ago and then lost track of it. The termination story is really a memory-management story.
Reality
Luna documented the problem, warned the worker and arranged training, then lost track of the attendance policy it wrote itself. Humans supplied the prompt and signed the termination.
Reality
Andon Labs says its agent Luna flagged an employee late for 17 of 23 shifts, but only after being prompted to search its memory. The judgment held. The follow-through did not.
Reality