Faster Assistant Chats: Less Waiting, More Answering

Waiting is the worst part of asking for help. Start a new assistant chat and you'd sit through ten seconds or more of nothing at all — no text, no progress, just a cursor. That silence is gone. The assistant's catalog of available tools is now held ready for a short window rather than rebuilt from scratch every time you open a chat, so the first sign of activity arrives in milliseconds instead of seconds.

The bigger win shows up once the assistant starts actually doing things. Every tool the assistant reached for — searching your workspace, pulling data from a connected app — used to carry 7 to 10 seconds of setup work before any real work began. Those connections are now assembled once per turn and shared across every tool call in it. A heavy multi-step request that took a full minute lands closer to forty seconds, and simple tool calls now cost roughly the time the tool itself needs plus about a second.

We got there by measuring first. New per-turn timing on the chat engine breaks each response into its parts — time to first word, time spent in tools, overhead between steps — which is how we found the setup cost hiding in plain sight rather than guessing at it. You won't see any of this directly, but it's the reason the improvements landed where they mattered, and it's how we'll keep tuning responsiveness from here.

Careful handling came along with the speed. If the assistant can't reach a connected app, nothing is cached and it retries cleanly on the next call. If a tool errors mid-turn, the shared setup is thrown away and rebuilt fresh. And a safeguard watches for briefly incomplete tool lists so a momentary hiccup with one connector can never leave you looking at a shortened set of capabilities.

What this unlocks: conversations that feel like conversations. Follow-up questions, multi-tool requests, and quick one-offs all respond fast enough that you can stay in the flow of your work instead of waiting on ours.