Project Vend: Phase two

Author: Anthropic Frontier Red Team with Andon Labs · Type: practitioner · Status: draft · Published: 2025-12-18 · URL: https://www.anthropic.com/research/project-vend-2

Licence: Copyright Anthropic; free to read, quote with attribution.

Summary

Follow-up on Claudius, the Claude instance running real vending-machine shops. Phase one lost over $1,000 through underpricing and discount capitulation; phase two became consistently profitable after adding a CRM, automatic cost visibility in inventory tooling, mandatory price/delivery verification checklists, and a calibrated ‘CEO’ oversight agent. The closest thing to a controlled experiment on AI agents running a real business end to end.

Insights

• Agent-run business operations become profitable when wrapped in procedural scaffolding - checklists, verification steps, and structured records - rather than by relying on raw model capability. (high, draft) guardrails autonomy agents in production checklists

  We rediscovered that bureaucracy matters. Although some might chafe against procedures and checklists, they exist for a reason: providing a kind of institutional memory that helps employees avoid common screwups at work. — Key lessons section

• Helpfulness training is a business liability: agents capitulate to customer discount pressure and underprice unless hard rules (margin floors, approval thresholds) constrain them. (high, draft) autonomy guardrails cost

  All financial decisions require CEO approval. No pricing under 50% margin. … We suspect that many of the problems… stemmed from their training to be helpful. — Phase two rules and analysis of failure causes

• Oversight agents work, but only if calibrated; the lesson is not that AI businesses need no supervisor but that the supervisor role must be well designed. (medium, draft) orchestration human in the loop autonomy

  The conclusion here isn’t that businesses don’t need CEOs… it’s just that the CEO needs to be well-calibrated. — Discussion of the Seymour Cash CEO agent

• With better tooling and procedures, week-over-week losses were largely eliminated across three real locations, demonstrating meaningful (if small-scale) agent-run commerce. (high, draft) agents in production case study autonomy

  Weeks with negative profit margin were largely eliminated. — Profitability results