Amazon Bedrock AgentCore payments is now generally available: Enabling agents to transact safely and autonomously at scale
Amazon Web Services Chethan Shriyan
AWS just made AgentCore payments generally available, letting AI agents autonomously pay for APIs, content, and inference. Agents can now spend real money on their own, with spending caps built in to stop them going rogue.
Based on reporting by Amazon Web Services, Chethan Shriyan — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Amazon has quietly solved a problem that sounds small but isn't: what happens when an AI agent needs to pay for something. AgentCore payments, built with Coinbase and Stripe, moved from preview in May to general availability today, and it's aimed squarely at a world where agents chain together dozens of tools and services on the fly, many of which now charge by the call rather than by the subscription.
The mechanics are worth understanding because they reveal how cautious AWS is being here. Agents get funded through Coinbase or Stripe Privy stablecoin wallets, built for microtransactions that often run to a few cents. Developer credentials sit locked in AgentCore's Identity Secrets Manager, and the agent itself never sees them; instead, short-lived tokens do the signing. AWS added a Quick Create option for Coinbase so developers can provision credentials without leaving the console, though Stripe Privy still requires a trip to its own dashboard.
Protocol support has expanded too. At preview, only the x402 protocol was supported. Now AgentCore also handles the Machine Payment Protocol, co-authored by Stripe and Tempo, and it's added an "upto" spending-ceiling scheme within x402 alongside the existing fixed-price "exact" scheme. That distinction matters more than it sounds — a spending ceiling lets a merchant charge for whatever compute or tokens actually got consumed during a call, rather than forcing everything into a flat rate up front.
Because agents are non-deterministic — they can misread a response as permission to spend, or double-pay after a bad retry — AWS wrapped every transaction in a payment session with a hard cap on spend and an expiry time, checked deterministically at the infrastructure layer before anything gets signed. Observability logs flow into CloudWatch and AgentCore Observability's dashboards, tracking things like transaction success rate and average value.
Real usage is already showing up in odd corners. Anchor Browser uses it to unlock paywalled content for browser-automation customers, Cloudflare's Monetization Gateway lets site owners charge agents per request, SpreadX's Incarna product pays for inference through a router called BlockRun, and Travala has wired it into travel-booking MCP servers so an agent in Claude can book a hotel mid-conversation. Elsa AI and Heurist AI use it for financial research APIs. None of these are hypothetical pilot decks — they're integrations AWS is naming today.
My take — AI-written commentary, not fact-checked reporting
Letting software spend money without a human clicking confirm is exactly the kind of thing that sounds fine until it isn't, and the fact that AWS built session caps and expiry timers first, not as an afterthought, says they know it too. The interesting tension here is that AWS is betting the agentic economy needs guardrails baked into infrastructure rather than trust in the model's judgment — which is probably the correct call, given how casually these systems have been shown to misread a prompt. Watch the merchants, not the agents: whoever controls the paywalls and CDNs deciding what gets charged per-call is quietly building the toll booths of this new economy.
Read more about this at: Amazon Web Services