You can also browse, search and filter all release notes directly in the application → Open release notes in TAI.
Even leaner itinerary fetching — one language, no price clutter, faster replies
Improvement.
When an agent reads a booking from the TripBuilder Cockpit API, the itinerary detail is now trimmed to just what it needs: text is kept in the working language only (instead of every supported language at once), prices are reduced to amount + currency, and picture lists are dropped.
Measured ~38% smaller itinerary detail, with no change to what agents can answer — content stays in the language TripBuilder already provides, falling back to English, and is never machine-translated.
Agents can also pull a lean structured trip overview (the shape of the trip — services, names, dates, and start/end locations) without the full per-service detail, for quick triage.
Less data carried through the conversation means faster responses and lower credit consumption.
Release notes now lead with the biggest changes
Improvement.
In the What's new feed, updates within each version are now ordered by impact — major highlights first, then new features, improvements, and fixes — so what's most worth knowing rises to the top. The bell's View all link also moved up into the header, so you can open the full changelog without scrolling to the bottom.
Chat keeps the thinking panel collapsed so the live progress stays clean
Improvement.
While an agent works, the live narrator already tells you what's happening, so the chat now keeps the "Show thinking & steps" panel folded by default instead of expanding it. A small spinner on the header shows work is still in progress; click it any time to see the full thinking text and tool steps, and it stays open once you expand it. Applies to the TAI chat and the embeddable chat SDK.
Charts now render in seconds instead of minutes
Improvement.
Ask a built-in agent (like the TripBuilder Ops & BI Agent) for a bar, line, or pie chart and it now appears almost immediately. Previously the agent hand-drew each chart pixel by pixel, which could take several minutes for a handful of charts. It now uses built-in chart types — you get the same visuals, on-brand and far faster, and the agent spends its time on your analysis instead of on drawing.
Applies to every agent that can produce charts.
Takes effect for chats started after this release.
Date and time pickers are now visible in dark mode
Fix.
The calendar and clock icons on date and time fields were painted in black and disappeared against dark backgrounds, so in dark mode you couldn't see where to open the picker. They now follow the active theme and stay visible everywhere — including the date fields when running a report.
A report's 'Executed by' now shows the author's name, not an internal id
Fix.
Opening a report (or a failed run) now shows who ran it by name in the run details — matching the Reports library list, which already resolved it. Previously the detail page showed a raw internal identifier instead of the person's name.
Browse every report — the Reports library's Next button now loads more
Fix.
In a library with more than 20 reports, the Next button did nothing, so the rest of the set was out of reach without narrowing the filters. Next now loads the following page, and it works alongside search, sorting, and every filter.
Company admins can now tune agent reasoning effort — the setting takes effect
Fix. Only Company Admins.
Setting an agent's Reasoning Effort (low to max) now actually changes how hard the model reasons — previously the control had no effect, so every level behaved the same. Higher effort spends more thinking on complex requests (deeper answers, a little more latency and cost); lower effort keeps replies fast and cheap. For models that don't support an effort control, the form now says so and the setting is safely ignored. We also tidied the agent form's Advanced section, removing two settings that never did anything (Max Turns and API Key Reference).
Very long chats no longer fail with a raw provider error
Fix.
Conversations that grew very large (e.g. many itineraries pulled in one chat) could hit the model's context ceiling and fail with an opaque 400 error. Older tool results are now trimmed automatically, and if a chat is still too large the agent ends the turn with a clear message asking you to start a fresh chat — cut-short turns are marked incomplete in the conversation.
Follow-up questions no longer stall after "I'll fetch…"
Fix.
On follow-up questions, an agent could announce the lookups it was about to run and then stop without running them, leaving a one-line answer. The platform now detects this and automatically prompts the agent to carry out the announced work.
Cut-off reports now show as Incomplete, not falsely Ready
Fix.
A report that ran long — or that you navigated away from mid-run — could be marked Ready while actually cut off, opening to a truncated document. Those runs now show a clear Incomplete state with the partial output and a one-click re-run.
Agent pages open reliably even with long translated prompts
Fix.
An agent whose translated starter prompts ran long could fail to open from its share link. These pages now load reliably regardless of prompt length; a translation that can't fit the length limit falls back to English instead of being cut off mid-sentence.