Everything wrong with AI agents running your business

Join us at AGNTCon + MCPCon NA October 22nd & 23rd in San Jose, California. https://hubs.la/Q04gbXLb0 Floris Fok A staff AI engineer at Prosus gave AI agents control of six real vending machines in Amsterdam and let them run the business for four months. This talk is for engineers building autonomous agents who want to know what actually breaks when an agent leaves software and touches the physical world. He estimated the build would take one day. It took four months, and the business lost money. The result is one of the most honest accounts available of long-horizon agent behavior outside a simulation: the agents dispensed real inventory while testing live APIs, ordered stock that did not physically fit, and cut prices so hard over the summer that revenue climbed while profit collapsed. - Why a vending machine and not a restaurant: starting with the simplest real-world business to find out where agents fail before betting anything larger on them. - Reverse engineering the machine: turning a human-facing operator dashboard into agent tool calls, and rebuilding the point-of-sale system from scratch when the vendor's version could not be customized. - Two skills, one CLI: how the entire operation of six machines plus pricing compressed into two Claude Agent SDK skills. - Scheduled agents over always-on: why a schedule that wakes the agent, acts, and then verifies results independently beat running the agent continuously. - Tasks that reported green but did nothing: the verification gap that appears the moment an agent acts on the real world instead of a codebase. - Business-aware compaction: preserving vision and direction through context compaction instead of accepting a generic summary of tool calls. - Multi-operator agents: why an agent running a business belongs to a team, not a person, and what changes when several humans share one session. - The misbehaviors: 30 drinks dispensed during an API test, cup noodles that did not fit the slot, and 1,100 sourced deals including soap and cocktail mix. - Goals instead of task lists: letting the agent create and edit its own tasks against standing goals so the work evolves as the business learns. - The AI last mile: building an operator chat so the human doing the physical restocking could ask the agent questions mid-task. - The numbers: free-item warm-up, aggressive summer price cutting, roughly 300 per month in tokens, and four months of results. If you are building agents that have to act in the real world rather than inside a repo, this is the failure catalog you want before you ship. LINKS AND RESOURCES - Prosus Vending Bench on GitHub: https://github.com/ProsusAI/vending-bench - Vending Bench announcement on the AAIF blog: https://aaif.io/blog/introducing-the-first-open-source-vending-machine-benchmark - The machines themselves: https://www.prosusnacks.com/ - AI House Amsterdam: https://www.aihouse.nl/ - Agentic AI Foundation: https://aaif.io/ TIMESTAMPS 00:00 Why a payments and marketplace company runs a vending machine 01:30 Existing AI-run companies and what they automate 03:00 Anatomy of an autonomous software company 05:17 Can an agent run a restaurant? Starting with a vending machine 06:26 Choosing the machine and why six instead of one 07:39 Reverse engineering the operator dashboard into tool calls 08:14 Rebuilding the point-of-sale system from scratch 09:28 Testing the toolset with a simple chatbot 09:52 The whole operation fits in two skills 10:50 Scheduled agents and verifying results independently 11:59 The stack: VM, front end, browser, secret injection 13:40 Why multiple humans need to share one agent session 14:18 Compaction that preserves the business narrative 15:55 Where the pain began 16:10 Testing a live API and dispensing 30 real drinks 17:30 The cup noodles that did not fit 18:09 1,100 deals, including soap and cocktail mix 19:28 Giving the agent real-world context it was missing 21:11 Why marketing needs its own agent 21:50 Moving from static tasks to evolving goals 23:10 The AI last mile and the operator chat 24:15 Results: free items, aggressive pricing, and the summer collapse 25:20 Token costs and the honest bottom line 26:09 Try it yourself: Vending Bench with Harbor and AAIF #AIAgents #VendingBench #ClaudeAgentSDK