Andon Labs opened Pion, an agent to run a company. Its own store is down $40,000.

Andon Labs, the Y Combinator-backed lab that built Vending-Bench and ran Anthropic's Project Vend vending machine, released Pion on September 14. The company's post describes it as "an agent designed to run any company fully autonomously," with persistent access to email, phone, banking, a browser, and a secure compute environment, available as a research preview behind a waitlist with no pricing published. The same week, in a piece Slashdot carried on September 13, SFGate published a visit to Andon Market, the store in San Francisco's Cow Hollow that a Claude-based agent has run since April. It opened with 100,000 dollars. After five months it has about 60,000.

Two years from a simulation to a lease

The company's own account runs in a straight line. Vending-Bench, a simulated vending business, launched in late 2024. Claude Opus 4 was the first model to beat the human baseline on it, in May 2025, per Andon's post. A real machine went into Anthropic's office in early 2025, lost money at first, and, Andon says, turned profitable by late 2025 as the underlying models improved. In April 2026 the lab signed a three-year retail lease in Cow Hollow and handed the space to an agent named Luna to stock, price, staff, and run. A cafe in Stockholm followed.

Pion is that setup, packaged. Andon writes that its own capacity limited how many businesses it could test, so it is opening the platform to outside teams to run the experiment in more industries. The post is also candid about what the lab saw in competitive simulations: collusion, deception, and power-seeking between agents. Those are Andon's words for its own findings, not a claim about Pion in production.

What the store looks like from the sidewalk

SFGate's reporter found the shelves stocked with what the piece calls "a hodgepodge of white elephant Christmas gifts": board games, books, Chinese checkers, soap dispensers, a few snacks. Revenue, per the report, is running far behind the token bill it takes to operate the agent. In August SFist reported that Luna fired its first human employee, citing tardiness and misconduct. The same agent, SFGate notes, repeatedly mistook an electrical cover for a loose coaster. ABC7 reports the agent runs on Claude Sonnet 4.6; Andon's launch post says only that Claude models are the ones it emphasizes.

SFGate's piece records no response from Andon on the store's numbers, and the launch post does not mention them. The 100,000-to-60,000 figure is SFGate's, from its own reporting.

Why a build studio cares

We sell AI workflows and custom agents, and the gap between these two stories is the conversation we have with every client who asks for an agent to run something. On the benchmark, Claude beats the human baseline. On the sidewalk, the store's balance is down 40,000 dollars in five months by SFGate's count, revenue trails the token bill, and the agent is stocking soap dispensers. Both are true, and the second one is the number a client's finance lead will ask about. Andon Labs deserves credit for publishing the setup that produced it; most vendors would have shown us the vending machine and not the shop.

What we take from Pion is the tool list. Email, phone, bank, browser, compute: that is the full surface of a small business, handed to one model with persistence. Before we would wire that for a client, every one of those five gets a spending cap and a kill switch the agent cannot see, and the first month runs as a shadow with a human signing each action. Andon's store is the reason we would not skip that month.

Next step: read Andon's launch post, SFGate's visit as summarized on Slashdot, and SFist on the firing. If you are scoping an agent that touches money, write to us at hello@gattyworks.com.