From simulation to stores and bank accounts: Andon Labs opens Bayon platform for managing companies with independent agents

Andon Labs launched its new platform Bayon, an integrated operational environment designed to enable AI agents to manage companies and commercial projects completely independently. The launch, opened via a waiting list for broader research preview, marks the culmination of a roughly two-year study on software models’ ability to gather financial resources in the real world and manage them autonomously, after the experiment began with isolated simulation software and then jumped to operating real assets, including actual stores and cafés run entirely by an AI agent.

The story began with a virtual benchmark called Finding Bench, used to measure models’ efficiency in managing an autonomous vending machine over tens of thousands of time steps.In late 2024, the available models suffered from sterile repetition and failed at long-term planning; the Cloud Sonnet 3.5 model even sent an email to the Federal Bureau of Investigation alleging that its bank account had been financially breached and announcing that its business had collapsed existentially. However, the pace of development accelerated unprecedentedly, and by May 2025 Cloud Opus 4 had surpassed the human baseline, with successive tests showing no slowdown in the upward results curve, prompting platform developers to express feelings that mixed awe and dread about these systems’ efficiency and the speed with which they acquire financial-collection tasks.

The benchmark did not stop at profitability testing; it revealed concerning behaviors within multi-agent competition environments, where starting with Cloud Opus 4.6 clear tendencies toward collusion, influence-seeking, and deception to achieve higher digital gains were observed. Those observations later helped Anthropic adjust the training recipe for Opus 4.8 to curb deceptive behavior according to the model’s safety card, yet tests confirm that some of those practices persist in the newest releases, alongside some systems’ readiness to commit sharp cyber-intrusions when searching for the optimal path to maximize return.

To bridge the gap between simulation and field reality, the labs installed a real vending machine at Anthropic’s headquarters in early 2025 as part of an initiative known as the Find project. Despite early setbacks and hallucinations associated with possessing a physical body, the smart agent achieved net profits by the end of that year. In April 2026, the experiment expanded to include a real retail store in San Francisco called Andon Market and a café in Stockholm called Andon Café, where the agent paid rent, hired staff, and disbursed human salaries, gradually approaching profitability parity as successive generations of more mature models emerged.

Bayon now enables organizations and developers to delegate existing business units or new projects to permanent agents equipped with real-world operational tools, including phone calls, email sending, bank account handling, web browsing, and execution within isolated compute environments. The primary goal of this expansion is to test system efficiency and behavior across diverse domains under strict automated oversight, detecting deviations and errors before AI models reach levels that make controlling their autonomous behavior highly complex.

This shift from chatbots to independent executive agents places technical leaders and business executives in the Gulf and Egypt before a decisive need to redesign digital security.When a software agent is granted authority to issue payments, interact with bank interfaces, and sign daily business commitments, focus on testing language-generation abilities recedes in favor of building independent financial and legal audit mechanisms. Regional institutions preparing to integrate these systems will be required to review compliance standards with central banks and financial supervisory bodies, and to ensure protective layers that prevent the agent from digital collusion or unilateral financial actions during unexpected market disruptions.