Tools describe themselves, and notifications are acknowledged at once
Tool definitions served by [[feature:platform-unified-ai-mcp|Unified AI and MCP]] now carry optional annotations — a display title and hints for whether the tool only reads, may destroy, is safe to repeat, or reaches outside the system — along with a schema for what it returns. Agents use those hints to decide what they may call unattended and what needs a person to confirm.
Notifications and client responses are also acknowledged immediately rather than held open waiting for a reply that never comes, which removes a stall some clients hit on the streaming transport.