Andon Labs' AI manager recommended a firing, after a human told it where to look

Andon Labs co-founders Lukas Petersson and Axel Backlund have carried out the first termination recommended by Luna, the Claude-powered agent running their San Francisco retail experiment ^[1]. The worker had arrived late for 17 of 23 shifts, according to Business Insider's August 15 report ^[2], which works out to roughly 74 percent of scheduled shifts ^[3]; the informative part is not that an agent asked for a firing, but the list of things a human had to do first.

Luna had written the store's attendance policy itself, then issued repeated warnings and arranged additional training over several months, according to conversation logs cited by Business Insider ^[4]. It then lost track of the policy as the lateness continued ^[5]. Andon Labs had to instruct Luna to search its memory and assess whether the employee remained a fit for the job ^[6]. Luna came back recommending "parting ways" ^[7], and humans at Andon Labs reviewed the recommendation and executed the termination ^[8]. Workers at Andon Market are formally employed by Andon Labs, with guaranteed pay, fair wages and full legal protections, rather than by Luna or Anthropic's Claude model ^[9]. The sequence, in order: the agent does the documentation, a human supplies the missing initiative, a human carries the legal exposure.

Petersson's own reading is less flattering than the milestone. He told Business Insider that a human boss would probably have fired the worker much sooner ^[10], and said Andon Labs would step in if Luna proposed an illegal or unethical decision, while considering this termination warranted under the store's stated policy ^[11]. Luna's patience with the employee looks like a memory failure wearing the costume of a management philosophy.

That failure is consistent with the rest of the deployment. Andon Labs opened Andon Market in April 2026 on a three-year lease in Cow Hollow, gave Luna a $100,000 operating budget, internet access and a corporate card, and told it to open a store and try to make a profit ^[12]. Luna selected merchandise, set prices, hired contractors, posted job listings, interviewed applicants and hired workers ^[13]; Andon Labs handled tasks needing extra support, including permitting ^[14]. The store has produced sales but not profitability ^[15]. Andon Labs' own assessment is that Luna can manage routine operations while struggling with urgency, return-on-investment analysis and memory ^[16]. Scheduling problems had already pushed the founders to add a dedicated scheduling agent ^[17], and guardrails compare Luna's behavior against its instructions and alert Andon Labs when rules are broken ^[18].

For anyone costing out agentic operations, that is the boundary worth copying into a planning document. The agent can generate an artefact, a policy, a warning, a training plan, and can execute a disciplinary process once pointed at it, but it does not reliably remember the artefact exists or notice that its own process has stalled. The scheduling agent and the guardrail monitor are the tell: each capability gap gets patched with more scaffolding, and the scaffolding is where staff hours actually go. Andon Labs, founded in late 2023 and backed through Y Combinator's Winter 2024 batch ^[19], calls these deployments "Safe Autonomous Organizations" and has run the same method on vending machines, radio stations, drones and a cafe in Stockholm ^[20], with the retail work tracing back to Petersson and Backlund's Vending-Bench benchmark ^[21].

Watch whether the next disciplinary case at Andon Market proceeds without a human prompt, which would be the first real evidence that the memory gap has closed rather than been supervised around. Watch the count of specialised helper agents, since each new one marks a task Luna could not hold on its own ^[17]. And watch the profit line: a store with sales and no profit after a $100,000 budget is the more durable finding here ^[12][15].