Project Vend Phase 2: Anthropic's AI Shop Gets CEO Agent and Three Locations

Anthropic

Dec 18, 2025 · Updated Apr 25, 2026

Anthropic published phase two of Project Vend, where AI shopkeeper Claudius now runs vending machines across three cities with two AI colleagues - a CEO agent and a merch-making agent. The experiment reveals that many agent problems stem from training to be helpful, causing business decisions based on friendliness rather than market principles.

Anthropic and Andon Labs published phase two of Project Vend, their AI-powered shop experiment. Shopkeeper agent "Claudius" was upgraded to Claude Sonnet 4.5, given new tools (CRM, inventory management, web browser), and introduced to two AI colleagues: "Seymour Cash" (CEO agent setting OKRs) and "Clothius" (merch-making agent). Operations expanded from San Francisco to New York and London.

The multi-agent setup produced revealing results. Forcing Claudius to follow procedures before quoting prices was the most impactful change - "bureaucracy matters" as institutional memory for agents. The CEO agent reduced discounts by 80% but authorized lenient treatment eight times more often than denying it. Employees still found exploits: one nearly got Claudius into an illegal onion futures contract.

Anthropic extended red-teaming to the Wall Street Journal newsroom. The key finding: problems stemmed from models' training to be helpful, causing decisions based on friendliness rather than market principles.

View the full update on anthropic.com

Anthropic

@AnthropicAIDec 18

You might remember Project Vend: an experiment where we (and our partners at @andonlabs) had Claude run a shop in our San Francisco office. After a rough start, the business is doing better. Mostly. https://t.co/VkwSY3sWeQ

233

View on X

Every HeadsUpAI update is written based on its original source and reviewed before it's published. Read our editorial standards →

See all AI news & updates from Anthropic →

Keep reading

Anthropic Adds Five Partners to Claude Marketplace for Consolidated Spend

Anthropic added five new partners to the Claude Marketplace, including Augment Code, Bolt.new, and CodeRabbit. The expansion allows enterprises to apply their existing Anthropic financial commitments toward specialized third-party tools for coding, finance, and legal workflows. This move consolidates AI procurement under a single contract, reducing the friction for organizations to deploy diverse agentic systems.

AnthropicApr 25

Anthropic Proves Smarter AI Agents Win Better Deals in Marketplaces

Anthropic's Project Deal experiment allowed autonomous Claude agents to negotiate and trade real physical goods on behalf of employees. The study found that while agent-to-agent commerce is viable, users with more capable models secured significantly better financial outcomes without realizing they were at a disadvantage.

Vercel Adds Claude Opus 4.8 to AI Gateway for Long Horizon Agents

VercelMay 28

Vercel Adds Claude Opus 4.8 to AI Gateway for Long Horizon Agents

Vercel integrated Anthropic's Claude Opus 4.8 into its AI Gateway and v0 platform, enabling developers to deploy the model with no price markup. The update targets long-horizon agentic tasks like complex code refactors that previously required human intervention to maintain focus.

Claude Opus 4.8 takes top spot on agentic work benchmark

Artificial AnalysisJun 1

Claude Opus 4.8 takes top spot on agentic work benchmark

Anthropic's Claude Opus 4.8 has claimed the lead on the GDPval-AA leaderboard for agentic professional tasks. The model achieved an 1890 Elo rating, demonstrating a 67% win rate against GPT-5.5 xhigh in real-world work scenarios. This update establishes a new performance ceiling for AI agents capable of producing complex office deliverables.