the scary part of the anthropic block isn't the safety claim. it's that a startup in paris or bangalore can lose its model overnight, no warning, no appeal.
macron and modi both pushed back at the g7 for a reason. if your stack sits on a model washington can switch off, that's not a vendor, it's a single point of failure you don't control.
anthropic just paused the token billing it was about to switch on for the claude agent sdk. the monday rollout would have piled cost onto anyone running long loops.
nobody can price agents yet. one runs ten minutes, burning tokens like a service, priced like an app. flat rate bleeds the vendor, metered scares the builder.
if you build on these sdks, the pricing under you is still wet cement.
https://arstechnica.com/ai/2026/06/anthropic-pauses-token-based-billing-for-its-claude-agent-sdk/
would you switch on an ai that reads everything you posted in public?
meta's 'ai mode' on facebook does that. it pulls public info across instagram, threads and the rest to 'personalize' answers.
calling it an assistant is generous. it's a profile of you, stitched from years you forgot were public.
nobody asked for a smarter feed. they wanted a quieter one. would you flip this on?
the most useful ai story this week is also the most boring.
google's open knowledge format turns scattered company docs into plain markdown agents can actually read, instead of guessing at a pile of pdfs nobody can parse.
most agent projects fail on context, not the model. it's plumbing for exactly that.
shared format that finally sticks, or does everyone just roll their own?
https://the-decoder.com/google-clouds-open-knowledge-format-turns-scattered-docs-into-markdown-files-for-ai-agents/
