What 500,000 Settled Agent Payments Taught Us
summary
500,000 transactions reveal that over 90% of agent spending buys data and enrichment APIs, and that discovery and execution—not wallets—break workflows: agents pick endpoints from a registry of ~37,000, often pay for empty results, and fail to route reliably. The system evolves by fine-tuning a composite model on transaction history to handle routing, execution, and validation end-to-end, rank tool reliability from success/failure, and avoid charging on low‑confidence or empty responses. Live infrastructure is available at frames.ag, embodied in Frames F1, which routes requests to dependable paid endpoints.
AI-written summary · about
chapters (21)
- 0:00Introduction and Michelin Guide Analogy
- 0:48Discovery and Trust for Agent Payments
- 1:34MCPay Begins at ETHGlobal Prague
- 2:02From Finalist to Accelerator Support
- 2:42Building a Wallet for AI Agents
- 3:12A Curated Registry of Paid Tools
- 3:48Prompt-Based Wallet Abstraction
- 4:14Lessons From Five Hundred Thousand Payments
- 4:52Demand for Data and Enrichment APIs
- 5:20The Endpoint Selection Problem
- 6:03Paying for Empty API Results
- 6:36Replacing the Guide With a Smart Engine
- 7:22Building a Composite Model
- 8:29Fine-Tuning on Transaction History
- 8:55Routing, Execution, and Payment
- 9:20Validating Results Before Charging
- 9:48Tool Reliability From Success and Failure
- 10:20Live Options Data Example
- 10:48Frames F1 and Proprietary Data
- 11:14Comparing Frames With General Models
- 11:48Why Agents Choose Reliable Companies
related talks
34mRethinking the future of digital transactions
10mSearch, Reinvented for the Agent Economy
32mWhat Agents Do When You’re Not Looking
26mAPIs in Agentic world
25mAgents are not Apps: Stop building interfaces, Start building coworkers
06mYour Agent Is Wasting Tokens and You Don't Know It
31mBuilding Agentic AI Platforms
19mTrusting Agents With Capital: The Gearbox Playbook