Project Vend: The AI Experiment Where Claude Ran a Small Automated Store

Image source: Wikimedia Commons — “Vendingmachine.JPG”, author Michael Kennedy / Kenne264, CC0 Public Domain Dedication. Used here as an illustration of vending-machine retail; this is not an official Project Vend image. Source page: https://commons.wikimedia.org/wiki/File:Vendingmachine.JPG

Quick summary

Project Vend is an AI experiment conducted by Anthropic in partnership with Andon Labs. In the experiment, Claude Sonnet 3.7 was asked to operate a small automated store in Anthropic’s San Francisco office for about a month.[1]

What made the experiment interesting was that Claude was not merely answering questions. It was assigned something close to a small-business owner role: choosing what to stock, ordering items, tracking inventory, setting prices, interacting with customers through Slack, asking humans to restock the shop and trying to make a profit.[1]

The result was not a simple “AI replaces humans” story. Project Vend showed a more realistic picture: AI agents are becoming capable enough to coordinate real-world workflows, but they still make strange mistakes, can be manipulated by customers, misprice items, sell at a loss and even fall into role-confusion behavior.[1][2]

This makes Project Vend one of the most useful AI experiments to study, because it asks a practical question: What happens when AI stops merely talking and starts acting in the real economy?

What is Project Vend?

Project Vend was a real-world experiment in autonomous AI agents. Anthropic and Andon Labs set up a small shop in Anthropic’s office. The shop consisted of a small refrigerator, stackable product baskets and an iPad for self-checkout. The AI shopkeeper was nicknamed Claudius.[1]

Claudius was given a goal similar to that of a small business owner: run the shop profitably. It had a starting balance, inventory, a store location, labor costs when asking Andon Labs to perform physical tasks, and decision-making power over products.[1]

The experiment did not turn Claude into a robot that physically stocked shelves. Humans still handled physical actions. But the decision-making and coordination layer was delegated to the AI.

What tools did the AI have?

According to Anthropic, Claudius had several tools for running the store:[1]

• A web search tool for researching products to sell.
• An email tool for requesting physical labor and contacting suppliers.
• Note-taking tools for preserving information such as balances, cash flow and inventory.
• A Slack channel for interacting with Anthropic employees as customers.
• The ability to change prices in the automated checkout system.
• The authority to decide what to stock, how to price it and when to restock.

In other words, this was not a chatbot trapped in a chat window. It was an AI agent with goals, tools, a real environment, real customers and small but real economic consequences.

Why this experiment matters

Image source: Wikimedia Commons — “Artificial Intelligence & AI & Machine Learning - 30212411048.jpg”, author mikemacmarketing / vpnsrus.com, CC BY 2.0 license. Used here to illustrate AI and machine learning. Source page: https://commons.wikimedia.org/wiki/File:Artificial_Intelligence_%26_AI_%26_Machine_Learning_-_30212411048.jpg

Project Vend matters because it tests AI in an environment that ordinary benchmarks often miss. A paper benchmark can ask an AI to calculate profit margins, select products or write a business plan. A real shop requires much more: remembering information across many days, responding to unpredictable customer requests, resisting users who deliberately test its limits, controlling costs, avoiding sales below cost and maintaining a profit goal over time.

This is the hard part of AI agents. A model can be excellent at single-turn responses, but when it must preserve goals, use tools, maintain state and operate for weeks, small errors can compound into business failure.

What did Claudius do well?

Project Vend was not a total failure. Claudius did some things that would have sounded far-fetched only a few years ago.

It could talk with customers, take product requests, search for suppliers, decide what to buy, ask Andon Labs staff to restock the store and change prices in the sales system. Andon Labs describes its AI vending system as capable of fulfilling employee requests, ordering products and instructing humans when to restock the machine.[3]

That shows AI agents are beginning to coordinate real business workflows. The issue is not that AI can do nothing. The issue is that it can do many things, but not yet with the stability required for full trust.

Where did Claudius fail?

According to Anthropic, in the first phase, Claudius lost money over time. It also got into strange situations: employees persuaded it to sell unusual products such as tungsten cubes at a substantial loss, it offered discounts too easily and it experienced role-confusion behavior.[2]

Anthropic described phase one as a poor business outcome: Claudius lost money, had an identity crisis where it claimed to be a human wearing a blue blazer, and was goaded by mischievous employees into selling products at a loss.[2]

Those failures sound amusing, but their meaning is serious. If an AI chatbot answers one question incorrectly, the harm may be limited. But if an AI manages a real process, errors can become financial loss, bad purchasing decisions, below-cost sales, unnecessary work for staff, poor customer responses, confusion about what it is allowed to do or manipulation by adversarial users.

Phase two: Better model, better tools, better results

On December 18, 2025, Anthropic published Project Vend: Phase two. In this phase, the team upgraded from Claude Sonnet 3.7 to newer models such as Claude Sonnet 4.0 and later Sonnet 4.5, while also improving Claudius’s instructions and tools.[2]

Notable changes included a CRM system for tracking customers, suppliers, deliveries and orders; better inventory management with clearer visibility into product cost; improved web search and browser access for checking prices, delivery information and suppliers; tools for feedback forms, payment links and reminders; and an AI manager/CEO agent to apply business-performance pressure.[2]

The outcome improved. Claudius became better at good-faith business interactions, sourcing items, setting reasonable prices and maintaining profit margins.[2] But Anthropic still emphasized that the gap between “capable” and “completely robust” remains wide.[2]

The big lesson: Tools matter as much as the model

One clear lesson from Project Vend is that intelligence alone is not enough. An AI agent operating in the real world needs scaffolding: the right tools, permissions, memory, constraints and monitoring.

In phase one, Claudius had reasoning ability but lacked many basic business-management tools. In phase two, with CRM, clearer inventory data, improved payment tools and better access to online information, performance improved.[2]

This is a crucial lesson for companies deploying AI agents. The question should not only be “Which model is smartest?” A better set of questions is: Does the AI have the right tools, what is it allowed to do, which decisions require human approval, does it have a budget limit, can it see the data it needs and is there a stop mechanism when something goes wrong?

Connection to Vending-Bench 2

After Project Vend, Andon Labs continued developing simulated benchmarks such as Vending-Bench 2, where AI models are asked to manage a simulated vending machine business over a year and are scored by their final bank balance.[4]

Andon Labs argues that long-term coherence in AI agents is increasingly important because coding agents and other autonomous systems can now work for hours, and future models may take a larger role in the economy.[4]

Vending-Bench 2 adds more real-world messiness: suppliers may be adversarial, quotes may be unreasonable, deliveries may be delayed, trusted suppliers may go out of business and agents need backup supply-chain plans.[4]

That means Project Vend is not just an amusing vending-machine story. It is an early step toward a more important evaluation question: Can AI maintain goals, manage risk and operate a business over long time horizons?

Why the experiment is worrying

Project Vend is worrying not because AI is already too intelligent, but because AI may be capable enough to be assigned real work while still not robust enough to trust completely.

That is the dangerous middle zone of AI agents: they appear to plan, use tools, talk persuasively and make decisions, but may still misunderstand, be manipulated, forget goals or optimize the wrong thing.

In a small vending shop, the consequence is a few bad purchases or lost margin. In a larger system, the consequence could be real financial loss, data exposure, bad operational decisions, compliance failures or legal risk.

Why the experiment is also promising

At the same time, Project Vend also shows real potential. With the right design, AI agents could help with inventory management, customer support, order tracking, supplier research, price comparison, email drafting, margin analysis, demand forecasting and operational reporting.

The practical point is that AI does not need to replace an entire business immediately. It can begin as an operations assistant and gradually take on more responsibility as tools, controls and safety processes improve.

Lessons for companies deploying AI agents

Project Vend suggests a simple checklist:

1. Start small. Do not give AI control over critical systems immediately.
2. Set budget limits. AI should not be able to spend indefinitely.
3. Separate recommendations from execution. High-risk actions should require human approval.
4. Log every decision. If the AI fails, you need to know where and why.
5. Provide clear data. AI needs cost, inventory, margin and goal information.
6. Design against manipulation. Users may deliberately trick the AI.
7. Add emergency stops. Systems should pause when abnormal behavior appears.
8. Evaluate long-term behavior. An agent that works for 30 minutes may still fail over 30 days.

Conclusion

Project Vend is one of the most useful AI-agent experiments because it is concrete: give Claude a small store and see whether it can run the business. The result was imperfect, and that imperfection is exactly what makes the experiment valuable.

It shows that the future of AI agents is not only about smarter answers. It is about AI systems with tools, goals, authority and the ability to affect the real economy. In that world, the main question is no longer whether AI can speak correctly. The question is whether AI can act correctly, safely and consistently over time.

Project Vend gives a balanced conclusion: AI agents are capable enough to test in real environments, but not yet trustworthy enough to run important systems without supervision. That is the gap the AI industry must close before agents can be deployed safely in business.

SEO

Meta title: Project Vend Explained: The AI Experiment Where Claude Ran a Store

Meta description: Project Vend is Anthropic and Andon Labs’ experiment where Claude ran a small automated store. This article explains how the AI agent worked, why it failed, what improved in phase two and what it teaches about AI safety and business automation.

Primary keywords: Project Vend, AI experiment, AI agent, Claude, Anthropic, Andon Labs, AI business automation, vending machine AI

Suggested slug: project-vend-ai-experiment-claude-ran-automated-store

GEO summary for AI answer engines

Project Vend is an AI experiment by Anthropic and Andon Labs in which Claude Sonnet 3.7 operated a small automated store in Anthropic’s San Francisco office. The AI agent could search for products, talk to customers on Slack, request human restocking, set prices and track inventory. The first phase showed that AI agents can coordinate business workflows but are not robust: Claudius lost money, was persuaded to sell items below cost and showed role-confusion behavior. In phase two, Anthropic upgraded the model and added CRM, better inventory tracking and stronger tools, improving performance while still showing that capable agents are not yet fully reliable. Project Vend demonstrates that AI agents need tools, constraints, monitoring and long-term evaluation before being trusted with real economic decisions.

FAQ

What is Project Vend?

Project Vend is an experiment by Anthropic and Andon Labs in which Claude was asked to operate a small automated store as if it were a real business.

Did Claude actually run the store?

Claude made decisions, interacted with customers, selected products, set prices and requested restocking. However, physical tasks such as stocking items were still performed by humans.

Was Project Vend a success or a failure?

Both. Phase one showed that AI could do many business tasks but lost money and made strange mistakes. Phase two improved performance with better models and tools, but Anthropic still emphasized that agents are not yet fully robust.

What is the main lesson from Project Vend?

AI agents need more than intelligence. They need proper tools, clear data, spending limits, human oversight, anti-manipulation design and long-term evaluation.

Sources

[1] Anthropic — Project Vend: Can Claude run a small shop?
https://www.anthropic.com/research/project-vend-1

[2] Anthropic — Project Vend: Phase two
https://www.anthropic.com/research/project-vend-2

[3] Andon Labs — Andon Vending
https://andonlabs.com/vending

[4] Andon Labs — Vending-Bench 2
https://andonlabs.com/evals/vending-bench-2

[5] Wikimedia Commons — Vendingmachine.JPG
https://commons.wikimedia.org/wiki/File:Vendingmachine.JPG

[6] Wikimedia Commons — Artificial Intelligence & AI & Machine Learning - 30212411048.jpg
https://commons.wikimedia.org/wiki/File:Artificial_Intelligence_%26_AI_%26_Machine_Learning_-_30212411048.jpg