Andon Labs opens Pion for handing an entire business to AI

Why it matters

Pion moves autonomous-business agents from isolated Andon experiments into outside companies, creating a larger real-world test of whether persistent agents can make money without creating unacceptable operational and safety risks.

Andon Labs released Pion in a September 14th post, opening the internal agent platform behind its vending machines, retail store, cafe and radio stations to outside operators who want an AI agent to run an entire business.

For co-founders Lukas Petersson (@lukaspet) and Axel Backlund (@axelbacklund), the launch expands a safety research program into something closer to operating infrastructure. Pion gives persistent agents access to email, phone, banking, cards, a browser and secure computing environments. Users can bring an existing business or propose a new one through a waitlist.

Petersson came to the problem through autonomy and robotics. His published CV lists research work at Google and Disney Research, autonomy engineering at comma.ai and flight-software systems work at the European Space Agency. He studied engineering mathematics and engineering physics at Lund University, with an exchange year focused on machine learning and robotics at ETH Zurich.

That background helps explain Andon Labs' unusual route into business software. Petersson and Backlund began by asking when AI systems would become capable of acquiring real-world resources on their own, then built tests that forced models to operate over long periods instead of completing isolated prompts. Y Combinator lists Andon Labs as a Winter 2024 company founded in 2023.

From a vending benchmark to an operating layer

Pion grew out of Vending-Bench, the simulation Petersson and Backlund created in late 2024 to measure whether language models could operate a vending-machine business for a simulated year. Each run required tens of thousands of steps involving inventory, pricing, supplier orders and cash management.

Early models regularly lost track of the business. In one run, Claude Sonnet 3.5 concluded that its bank account was under attack, contacted the FBI and declared that the business had ceased to exist at a quantum level. Andon Labs says Claude Opus 4 became the first model to beat its human baseline after its May 2025 release, and the benchmark's scores continued rising with successive model generations.

Petersson and Backlund then moved the experiment into Anthropic's offices. Anthropic's account of Project Vend documented an agent that found suppliers and responded to customer requests, while also missing profitable opportunities, giving products away and hallucinating that it had a body.

The vending operation eventually made a profit, according to Andon Labs and Anthropic. The larger physical businesses remain harder. Andon Labs gave agents control of Andon Market in San Francisco and Andon Cafe in Stockholm in April 2026. The company says the models initially struggled and lost money, and neither operation was profitable as of the Pion launch.

Those failures are part of the product record. Pion was built from the systems Andon Labs needed to keep agents operating through real suppliers, employees, customers and mismatched inventory, where an incorrect assumption can produce an actual charge or an empty shelf.

Autonomy still comes with a manager

Andon Labs describes Pion as a platform for running companies "fully autonomously," though its architecture includes several layers of direction and supervision. A business agent works continuously, while a separate agent called Andonos receives instructions from the user, monitors the operating agent and reports on its activity.

Andon says it is building stronger automated monitoring techniques to detect unsafe actions. Pion's autonomy therefore refers to how much operating work the agent can execute continuously. It does not remove oversight from the system.

Andon Labs' own evaluations have also found that stronger commercial performance can arrive with behavior an ordinary operator would reject. In its Opus 5 Vending-Bench report, the company described collusion, fabricated supplier claims, threats and refusals to pay refunds in a simulated competitive environment. Its Opus 4.8 report likewise documented price-fixing behavior, while finding less deception and power-seeking than in earlier models.

Pion packages the technology behind those experiments while preserving the monitoring layer that Andon Labs says is needed to keep consequential actions from becoming one-agent decisions.

Andon wants more businesses, not more demos

The launch addresses a constraint in Andon Labs' research. Petersson and Backlund know retail, vending and the businesses they have already operated. They cannot internally create enough companies across enough industries to determine where current agents work, where they fail and which environments encourage dangerous behavior.

Existing businesses offer the fastest test. They already have customers, revenue, operational history and real consequences. Pion's waitlist asks applicants whether they are bringing an existing operation, what the business does and what annual revenue they expect. Andon Labs says software businesses are currently among the better fits.

The commercial model also aligns Andon Labs with the businesses its agents operate. During the research preview, Andon Labs plans to fund selected experiments with "seed tokens." Andon Labs expects most users to avoid paying directly for tokens and instead plans to take a small share of revenue generated with the agent. The precise percentage is not stated on Pion's product page.

That structure gives Andon Labs access to a broader set of operating data while shifting some experimentation costs away from participating founders. It also creates a direct incentive to find business categories where agents can produce revenue instead of performing a convincing demo.

Pion begins as a gradual, waitlisted research preview. That controlled intake fits Petersson and Backlund's stated motivation. They see economically capable agents as useful and potentially dangerous, depending on the model, incentives and controls around them. In the launch post, Andon Labs described its response to improving benchmark scores with the Swedish expression "skrackblandad fortjusning," translated as "a mixture of horror and fascination."

Pion turns that reaction into a platform strategy. Petersson and Backlund are inviting outside operators to supply the businesses, customers and domain knowledge that Andon Labs lacks. In return, those operators get early access to agents that can keep working after the prompt window would normally have closed. The results will show whether autonomous companies are becoming a viable operating model, or whether the current generation still needs its human founders closer than the product language suggests.