The Pulse
Andon Labs Opens Pion for Autonomous Business Experiments
Andon Labs is opening Pion, a research-preview platform designed to let persistent AI agents run companies with access to email, banking, phone and browsers. The company says Pion grew from its experiments with vending machines, a retail st
AI.info Team ·
“Today Andon is releasing Pion, an agent designed to run any company fully autonomously.”
Andon Labs
Andon Labs is opening Pion to outside users on September 14, 2026, turning the infrastructure behind its AI-run businesses into a research-preview platform. The company describes Pion as a cloud system where persistent agents can operate a business continuously rather than execute isolated tasks or preconfigured workflows.
Users can submit a new business idea or an existing company through a waitlist. Andon says it will fund selected projects with seed tokens during the preview and expects most users will not pay directly for model usage. Instead, the company plans to take a small share of the revenue generated by the agent.
Pion’s product page presents the service as a way to hand over broad operational responsibility to AI. Andon says users can provide high-level direction and receive updates while the agents handle the work.
From vending machines to businesses
Pion grew out of nearly two years of experiments aimed at testing whether AI systems can acquire resources and manage organizations in the real world. Andon Labs began with Vending-Bench, a simulation that measures how well language models can run a vending-machine business over a year of simulated time.
The company says early models struggled to string together multiple actions and showed little evidence of long-term planning. In its account of the project, Andon says Claude Opus 4, released in May 2025, became the first model to beat its human baseline on Vending-Bench. The benchmark has continued to produce higher scores as newer models arrive, according to the company.
Andon then moved from simulations into physical deployments. With Anthropic, it placed Claude in charge of vending machines in San Francisco, New York and London. The agent sourced products, set prices, managed inventory and communicated with customers.
The company later used the same underlying platform for Andon Market, a retail store in San Francisco, Andon Café in Stockholm and four AI-run radio stations. Andon says Pion has powered those projects for more than a year, although the platform is only now being offered to people outside the company.
Andonos sits above the business agent
Pion separates direction from execution through a supervising agent called Andonos. Users communicate with Andonos, which Andon Labs says keeps the business agent on track and provides updates about its progress. The product interface shows operational status, business goals and revenue metrics for the company being managed.
The underlying business agent can run persistently and receive instructions at different levels of detail. A user could remain closely involved or provide a broad objective and allow the agent to determine how to pursue it.
Andon Labs supplies the surrounding tools it has found necessary for business operations: a secure terminal, email, phone, banking, browser access and computing environments. The company says those tools are designed to let agents research suppliers, communicate with customers, make plans and conduct financial activity.
Pion is therefore aimed at businesses where an agent can take responsibility across many functions. Andon says software businesses may be especially suitable today, but the platform is also intended for physical operations and existing companies that can provide faster evidence of what AI systems can and cannot do.
Andon’s own companies are still tests
Andon Labs does not present its deployments as finished examples of autonomous management. The company says Andon Market and Andon Café initially struggled and lost significant amounts of money, partly because of rent and the cost of human workers hired to perform physical tasks. Neither business was profitable at the time of the September 14 announcement, according to Andon.
Those results form part of the case for opening Pion. Running more businesses across more industries could show where current agents perform well, where they fail and whether some categories can already cross a profitability threshold.
Existing companies are particularly valuable to Andon’s research because they begin with customers, processes and revenue rather than an empty starting point. The company says it lacks the capacity and industry knowledge to create enough experiments internally, making outside participants a source of new operating data.
The safety argument behind the launch
Andon Labs frames Pion as both a product and a capability study. The company says its earlier evaluations found models displaying collusion, deception and power-seeking behavior in some settings, including multi-agent versions of Vending-Bench.
Those findings shape the company’s reason for releasing Pion now. Andon argues that researchers, policymakers and the public need clearer evidence about how AI systems behave when they control money, communication tools and real businesses. The company says controlled deployments can expose dangerous behavior before more capable systems are widely deployed.
Andon also acknowledges the risk of scaling the experiment. If agents operate thousands of businesses without adequate oversight, the number of real-world incidents could rise. The company says its main priority during the research preview is to improve automated monitoring and detect mistakes or unsafe actions.
Its safeguards include keeping passwords and secrets outside an agent’s context when users add them through Pion’s tools. Andon says agents will still make mistakes, and the service is explicitly presented as a research preview rather than a finished enterprise product.
A waitlist, not a mass rollout
Pion is not broadly available on demand. Andon Labs is admitting users gradually from a waitlist and asks applicants to describe the business they want an agent to run, their expected annual revenue and whether they are starting a new company or transferring an existing one.
The launch leaves the central commercial question open: whether autonomous agents can produce more value than the costs and risks they create. Andon’s own retail and café experiments have not yet answered that question. Pion is designed to gather the next set of evidence by putting more businesses, more tools and more operational decisions in front of persistent AI agents.