A capable agent workflow can discover relevant instructions without loading every installed skill into context, while keeping discovery distinct from permission to act. The release says search follows the agent’s current read permission; reading a selected skill does not grant permission to install or run it. Permission-aware search locates candidate installed skills, with a separate read step for full instructions; execution authorization remains distinct.
The release says eligible installed skills remain discoverable when prompt-size limits leave them out of the skill list. Teams can avoid loading a complete skill library into every prompt while retaining a discovery path. The release also adds MCP Apps beside chat: compatible apps can expose forms and workspace files, with visible approval controls for messages and file opens. Backups may now be stored on external storage or Cloudflare R2 and restored into a fresh staging directory.
These are documented product changes, not evidence of adoption, reliability in production, or measured context savings. The notes also caution that instruction search can be partial: for some long files, only a prefix is covered, so a missing match does not prove that no suitable skill exists. Whether later releases document wider instruction-search coverage or evidence of teams using discovery while preserving execution permission boundaries.