Ask HN: Has anyone built an AI agent that spends real money?
This is not a formal paper, but rather a practical request-for-help post about the feasibility of an "AI agent that can autonomously spend real money." Its core value lies in clearly exposing real-world bottlenecks such…
Summary
This is not a formal paper, but rather a practical request-for-help post about the feasibility of an "AI agent that can autonomously spend real money." Its core value lies in clearly exposing real-world bottlenecks such as payments, risk control, compliance, and platform blocking, showing that although card networks are optimistic about this direction, developer infrastructure is still clearly lacking.
Problem
- The problem to solve is: after receiving user authorization, enable an AI agent to autonomously complete the full loop of browsing products, selecting an order, and making a real payment.
- This matters because it is a key step from "can recommend" to "can execute," directly affecting the deployment of agent-based commerce, automated procurement, and more broadly executable AI assistants.
- The main current obstacles include card issuers not cooperating with individual developers, Stripe's 3D Secure restrictions on off-session payments, e-commerce sites blocking browser automation, and the legal risks of automation on major platforms.
Approach
- The author is building an MCP server that connects AI agents to payment rails such as Stripe, PayPal, and virtual cards.
- The intended mechanism is simple: the user provides payment credentials once, and afterward the agent independently handles "find products → make decisions → pay."
- This approach depends on existing payment rails and browser automation rather than a new payment protocol, so it is directly constrained by payment authentication, anti-fraud systems, and website policies.
- The post also cites Visa's Intelligent Commerce and Mastercard's Agent Pay as industry signals, indicating that card networks are paving the way for "agentic spending," but developer-ready tools are still immature.
Results
- No formal experiments, benchmark data, or quantitative results are provided; there are no figures for datasets, accuracy, success rate, or transaction conversion rate.
- The strongest concrete progress is that the author is "already building" an MCP payment service connecting Stripe, PayPal, and virtual cards, and has publicly shared the repository:
clawpay. - A key practical conclusion stated in the post is that Stripe requires 3D Secure for off-session payments, which makes it difficult for an agent to complete payment seamlessly within current e-commerce payment flows.
- Another concrete conclusion is that mainstream e-commerce platforms block browser automation, and Amazon v. Perplexity (described in the post as March 9) is viewed by the author as evidence that browser automation carries real legal risk.
- A positive industry-level signal is that Visa Intelligent Commerce and Mastercard Agent Pay have been launched, but the post provides no data on their availability, coverage, or developer adoption.
Link
Run your own research radar
Turn arXiv, Hacker News, OpenReview, Hugging Face Daily Papers, and RSS into local Markdown, Obsidian notes, Telegram digests, and a public site.