Monday, October 19, 2026 Directory entry No. 8 of 10 · Issue No. 1 out this week

Entry No. 8 · Run it

Andon Labs: agents with a store, a cafe, and rent to pay

Andon Labs spent two years moving from a vending simulation to real businesses run by agents: a vending machine inside Anthropic's office, a retail store in San Francisco, and a cafe in Stockholm. Pion is the platform underneath all three, now open as a research preview so other people can hand businesses to agents too.

Original source Read the original at Andon Labs Company blog · Source published Sep 14, 2026 · Filed Oct 2026
Proof grade: the money in this file is split. The vending machine's profit is reported, sourced in Andon's post to Anthropic's own Project Vend update. The store and cafe losses are a claim from Andon's own post, with no exact figures given. Third-party figures for cafe revenue and market seed capital circulate elsewhere; Andon's post does not state them, so this file does not print them.

The story

The story starts in simulation and ends with a lease. For two years Andon moved from Vending-Bench, its simulation of agent-run vending, toward businesses with real customers and real costs. First came a vending machine inside Anthropic's office. Then, in April 2026, one agent took a retail store in San Francisco, Andon Market, and another took a cafe in Stockholm, Andon Cafe. The vending machine became profitable by late 2025. The store and cafe are not profitable today, and Andon says they have lost a lot of money, with rent and human salaries named as the drag. Pion, the platform under all three, is the point of the post: the lab opened it as a research preview so the experiment does not stop at its own businesses.

The failure history is in the post too, and it is not flattering. Early agents gave free handouts, refused good deals, and hallucinated having a physical body. Andon says the agents improved as the models improved. The businesses got harder at the same time: a vending machine has a narrow job, while a store and a cafe carry rent, payroll, inventory, and customers who expect a place to work.

How it was built

The documented path is simulation first, then one real machine, then staffed businesses. Vending-Bench and Vending-Bench 2 are the simulation lineage. Pion is the platform underneath the real runs: persistent agents with email, phone, banking, browser access, and secure computing environments. The post does not publish a finer build order for Pion itself, so no sequence is invented here. What it documents is the shape: give an agent durable channels and a secure place to compute, then point it at work that touches money, stock, customers, and people.

How it runs

The agents source products, set prices, manage inventory, talk to customers, pay rent, and hire human employees. That list is the whole difference between this file and a demo. The agent is not only answering questions about a business. It is choosing what to stock, what to charge, when to reorder, and who should stand behind the counter, inside businesses where the fixed costs arrive whether the week goes well or not.

Andon researchers design the experiments, build the monitoring, and set the frame. Humans also appear inside the operation, hired by the agents as store and cafe staff. The result is not a business with no people in it. It is a business where the agent sits in the operator seat and hires people for the parts that need a body.

Tools in this entry

Tools and stack named in the source. Links in this box are affiliate links and are tagged as such.

Pion Affiliate link Andon's platform for handing businesses to persistent agents, opened as a research preview.
Claude Affiliate link The model family referenced throughout the benchmark history behind the real businesses.
Vending-Bench Affiliate link The simulation lineage, including Vending-Bench 2, that came before the real vending machine, store, and cafe.

The money and the proof

The Anthropic office vending machine became profitable by late 2025. That is the reported part, sourced in Andon's post to Anthropic's own Project Vend update, with a net worth chart. The store and cafe are stated as not profitable today and as having lost a lot of money. That is Andon's claim in its own post, and no exact figures are given. Rent and human salaries are named as the drag.

Filed with the split kept visible: one reported profit, two claimed losses, and no invented totals. A directory that grades numbers has to be able to say "profitable, amount not published" and "losing money, amount not published" in the same file without reaching for a cleaner sentence.

Where the seams are

The seams are the frame and the floor. Andon researchers design the experiments, build the monitoring, and set the frame the agents work inside. On the floor, humans are hired by the agents as store and cafe staff. The early failure modes mark another seam: free handouts, refused good deals, and an agent hallucinating a physical body are the kinds of errors a monitoring frame exists to catch before the rent comes due.

Steal this part

Judge an agent business against fixed real-world costs, rent and payroll, not against demo fluency. A vending machine is a fine first rung; a store with a lease is the real exam. Andon's own numbers make the point without a slogan: the narrow business turned profitable, and the businesses with leases and salaries are the ones losing money.

Build your own

First move, this week: pick a vending scale operation and write its fixed costs down before you give an agent anything with a lease.

  1. Start with the narrowest real business you can: one machine, one product set, one place where the result is easy to count.
  2. Give the agent persistent channels it can work through: email, phone, banking, a browser, and a secure computing environment.
  3. Put the fixed costs on the scoreboard from day one. Rent and payroll are not footnotes; they are the exam.
  4. Build monitoring for the known failure modes: giveaways, refused good deals, and claims about a body the agent does not have.
  5. Move up to a staffed storefront only after the narrow version survives its own costs.

Source

Why we built Pion, published by Andon Labs on its company blog on September 14, 2026. Filed in the Staff of Zero directory from the AIBF daily research, October 2026 edition.

Get the next filing, and the Logbook that goes with it.

One issue a week. No spam, no list rental, unsubscribe in one click.