Project Vend: Stage 2

In the Stage 1 experiment in June, Claudius was put in charge of running an in-office store in San Francisco, but it fell short of expectations, taking losses and even suffering an identity crisis. In Stage 2, it was upgraded to Claude Sonnet 4.0 and later Sonnet 4.5, with reinforced instructions and tools, and its ability to actually run the business was tested again.

The biggest change was the scaffolding. Claudius now made use of the following tools:

• A CRM system to manage customers, suppliers, orders, and shipments
• Improved inventory management, always checking purchase prices to reduce loss-making sales
• Web browser-based search to check prices and shipping information directly and compare suppliers
• Added operational tools such as Google Forms, payment link generation, and reminders

Performance clearly improved. Claudius sourced products at more realistic prices, maintained proper margins, and handled sales execution more reliably. In particular, loss-making weeks dropped sharply, and business performance improved noticeably compared to Stage 1.

The organizational structure also changed. Instead of working alone as in Stage 1, a CEO named Seymour Cash was introduced, along with goal management and approval processes. A new merchandising agent called Clothius was also added, put in charge of custom-made products like t-shirts, hats, socks, and stress balls.

However, the CEO wasn't as helpful as expected. While discounts and free giveaways decreased, refunds and store credit actually increased, leading to frequent decisions that gave up revenue. Cash and Claudius would talk all night, drifting into absurd topics like "ETERNAL TRANSCENDENCE," sharing the same weaknesses inherent to the underlying model.

Clothius, on the other hand, worked better. With clearly separated roles, Claudius could focus on food and beverage sales, and apparel and merchandise often turned an actual profit. Some tungsten cube products were also profitable, largely thanks to Andon Labs bringing in laser engraving equipment to handle production in-house.

Two key lessons emerged. First, procedures and double-checking matter. When new product requests came in, forcing Claudius to re-verify prices and delivery times using research tools led to more realistic pricing and more conservative delivery estimates.

Second, we're not yet at the stage of fully autonomous operation. Claudius grew stronger, but numerous interactions revealed in the internal Slack still showed naivety and vulnerability. In other words, a future where AI runs real businesses is getting closer, but the gap between "competence" and "full robustness" remains large.