Cached conversation context reuse, adaptive task step timing, and browser benchmarks
improved · reached everyone november 20, 2026 · in version 2026.11
Extended conversation turns reuse cached prompt history across planner steps to cut latency and cost, long multi-step tasks skip slow model calls when nearing time limits to keep moving, and browser tasks undergo continuous end-to-end benchmarks.
what changed
- Long conversations reuse cached prompt context across consecutive reasoning steps, reducing latency and work-unit usage.
- Multi-step tasks approaching time thresholds proceed directly to next actions rather than waiting on long model calls.
- Automated browser workflows undergo continuous end-to-end evaluation against real portal tasks before and after changes.
areas: Conversations · Everyday Experience