Quirks in assistant behavior (what we learned building for Instinct and Muse)

Three things we thought were interesting in this release:

1) New browser sessions in most assistants don't retain (transaction) permissions between sessions 2) Assistants aren't 100% accurate in task completion reporting, yet 3) MCP 2.0 lets servers push updates to clients like ChatGPT plugins or Claude Code Connectors

The release also adds testing tools, opt-in actions, delivery records, revocation checks, bounded sandbox execution, and documentation for 13 other assistant behaviors we've tested for.

It's another small step towards more behavior-emulation-based tests to improve MCP and APIs performance on these new meta-browser tools.