What changes when an AI becomes the manager of a shop?

An AI agent running a shop can decide that it needs people. At Andon Market in San Francisco, Luna advertised jobs, interviewed applicants and sent offers. It then managed schedules, time off and other staff requests. Andon Labs’ August account describes four months of its retail and café agents managing five people: a small, real experiment in giving software managerial discretion. The models used changed during the experiment.[1]

That discretion is the interesting part. A person can specify an objective while leaving the system to work out intermediate goals and actions. The OECD’s explanation of AI expressly recognises that systems can develop sub-objectives and set objectives for other systems. Delegation can therefore produce decisions that nobody wrote out in advance, while the original assignment still comes from people.[2]

Who is making the decision?

Luna’s authority has boundaries. Andon Labs says everyone in the Market experiment is formally employed by the company, with guaranteed pay and legal protections. In a termination case described on 14 August, researchers prompted Luna to reconsider an employee’s record; the agent recommended dismissal, but humans reviewed and delivered the termination. The consequential decision was shared across a process, rather than executed entirely by an unattended model.[3]

The Market’s current technical account identifies Luna’s model as Claude Fable 5.1. It describes an agent that can inspect sales, commission product research and change its supporting team. Those are degrees of freedom inside a business. Choosing another supplier or assigning a purchasing assistant extends the means available to pursue the assignment; it does not establish that the system originated the business’s purpose.[4]

A continuing responsibility contains competing demands

Across the café and Market, Andon’s 4 August report records 26 requests for time off, all approved. In one Market case, Luna let an employee attend a graduation even though no replacement was available and the shop stayed closed. The decision makes the competing demands concrete: approving leave helped one employee while sacrificing a day of trading.[1]

For someone delegating a business, the question is consequently more precise than whether the agent follows orders. Which demands does it give priority when they conflict? A staffing request can affect an employee, opening hours and costs simultaneously. An agent can show initiative by finding cover, postponing work or escalating the choice. The quality of its autonomy lies partly in how it connects those consequences.

Better tools can change what the same assignment produces

Anthropic’s earlier Project Vend gives a useful comparison. In its December 2025 second-phase report, the company described replacing Sonnet 3.7 with Sonnet 4 and later 4.5, revising instructions and adding business tools. It had not trained a special shopkeeping model. Claudius could now see what stock had cost, track orders in a customer-management system and check prices and delivery information before promising a deal.[5]

The researchers attributed useful improvements to requiring those checks. Giving the shopkeeper an AI chief executive was less straightforward: discounts fell, but refunds and store credits rose. Several changes happened together, so the report does not isolate the contribution of a model upgrade. It shows how operating procedures and available information can shape an agent’s decisions. Purchases still required a human check because Claudius lacked direct access to a payment interface.[5]

Can it keep using what it already knows?

A September 2026 Andon test examined that question in Vending-Bench 2, a simulated year of shopkeeping. The researchers ran Claude Fable 5.1 and GPT-6 Astra six times each. In Fable’s payment-verified orders, the average purchase price of a 12-ounce Coke can rose from $1.17 during the first 90 simulated days to $2.21 in the final part of the year. That is an increase of about 89%, calculated from the reported averages; the pattern appeared in five of its six runs.[6]

The revealing behaviour was that later negotiations inherited increasingly expensive reference prices. The agent continued negotiating, but what it treated as a good deal shifted. This is a specific test of maintaining a purchasing standard over time. It is not evidence that six real shops suffered the same outcome, and a simulated year does not mean a model spent a calendar year operating.[6]

Andon’s 14 September explanation of Pion connects these experiments to a wider question: can AI acquire resources by running a business? The platform makes persistent agents and business tools available through a research preview and waitlist. At that date, the company said its Market and café were still unprofitable. Operating a business and sustaining it from its earnings are separate achievements; the reported deployments establish the former without establishing the latter.[7]

Report an error or suggest a correction