Chat and AI Agents reliability improvements

Sun’s Out shipped on July 15. The surface has looked mostly the same since—underneath, Chat and AI Agents have changed a lot. This update covers ongoing improvements to citation quality, multi-turn context, and answer reliability across both.

Citations and groundedness improvements

• Citations from Channels deep-link straight to the source, the same way Docs citations always have
• Video citations fall back to a thumbnail or clip when the original can’t be shown
• When a citation comes from earlier in the conversation rather than a fresh search, it’s now labeled instead of dropped silently

More consistent follow-ups

• Chat holds on to search results, Channels data, and doc titles across a conversation, so follow-up questions get answers grounded in what came before
• Turns that include citations stay in the conversation history, reinforcing that citing sources is the default, not optional
• AI Agents now run on the same grounding logic as Chat, so both behave consistently

AI Agents that cite outbound

• Email answers use the full formatting engine, so citations survive the trip to your inbox
• Scheduled and triggered runs from AI Agents get an explicit reminder to include citations
• Creating a doc or a comment through AI Agents now returns a direct link to it

Fewer failed turns

• Chat can now take more search steps in a single conversation before hitting its limit, so it can fully answer longer, multi-step questions
• If Chat does hit its budget, it now gives you its best answer instead of failing outright
• Fixed a set of bugs causing intermittent errors when replaying conversation history
• Improved diagnostics when a prompt fails validation, so issues are easier to identify and fix

New measurements and quality tracking

• We now track citation health as a metric across Chat, Slack, and email
• We slice resolve rate by conversation depth and failure reason, catching regressions before they compound
• Developed multi-turn Chat evals to track regressions of quality over the span of a conversation

Ask a multi-step question in Chat or through an AI Agent and see the citations hold up throughout.