The framing buried in Cloudflare’s Monetization Gateway announcement is more consequential than the payment rails it ships with. Their argument: for 30 years the web ran on content-for-attention, monetized through ads and subscriptions — and that bargain breaks when the dominant reader is an agent that ignores ads, holds no subscriptions, reads a page once, takes what it needs, and leaves.

If that’s true, the unit of the web stops being the pageview and becomes the request. Cloudflare’s gateway lets you charge per call for pages, datasets, APIs, or MCP tools, with a single control plane for payment and access policy at the edge. The MCP-tool line is the one worth sitting with: the tools your agents call are about to have meters on them.

That turns an architecture decision into an ops problem I recognize from LLM cost governance. Right now most agent frameworks treat tool calls as free — retries, fallbacks, and speculative fan-out are cheap when the only cost is latency. Put a per-request price on every external tool and every retry storm becomes a line item. Suddenly your orchestration policy is your budget policy, and observability has to track spend per tool per run, not just tokens.

The HN thread is split on whether stablecoin settlement is the right substrate or an unnecessary crypto detour. I care less about the rail than the shape it forces: agentic systems built assuming free tool access will need a cost model the day their dependencies start charging.

My prediction: within a year, “cost per successful task” becomes a first-class agent eval metric, sitting right next to accuracy and latency. The teams treating tool calls as free are the ones who’ll get surprised by the bill.