You can also browse, search and filter all release notes directly in the application → Open release notes in TAI.

​

New Feature.
​

TAIDEMO-CAPETOWN-VIA-LONDON is a booked Zürich-Cape Town trip that changes planes at London Heathrow in both directions, with one traveler on a Swiss and one on an Indian passport. It is the first sample with a connection in a country the travelers never stay in, so you can see how transit and entry requirements are checked per traveler. Try it with Trip Guard or any agent equipped with the Sample Itineraries skill.
​

Learn more.

​

New Feature. Only Company Admins.
​

A connector can now reach an API that grants access through an OAuth 2.0 sign-in instead of a pasted token. Enter the provider's client settings, click Connect, approve access on the provider's site, and the connector keeps its token renewed automatically. If the provider withdraws access, the card shows it and company admins get an email to reconnect.
​

Learn more in Credentials.

​

New Feature. Only Company Admins.
​

Billing Center → Overview gains two tables — average, median, minimum and maximum credits per session, plus a session count — broken down by agent and by individual user. A total tells you how much was spent; this tells you what a typical session costs and how wide the spread is around it, so you can tell a one-off expensive conversation from the normal cost of the work. The median beside the average is the tell: far apart means a few sessions are pulling it up.
​

Personal agents are left out of the by-agent table, and the by-user table covers people — work done by an API key has no person to attribute it to.
​

The credit distribution tables by agent and by user in Billing Center. The numbers shown are fabricated demo data.

​

New Feature.
​

Every sample itinerary so far was built to surface a specific finding. Two new ones do the opposite: TAIDEMO-NEWYORK-BOOKED and TAIDEMO-NEWYORK-PLANNING are the same London-New York city break at booked and at quote stage, deliberately straightforward with nothing built in to go wrong — a baseline to read a flagged itinerary against, and a way to see how the checks change with booking stage. Try either with Itinerary Checker or any agent equipped with the Sample Itineraries skill.
​

Learn more.

​

Improvement. Only Company Admins.
​

Prompt Advisor runs used to start strictly oldest-first across every company, so a large burst of runs from one company could delay everyone else's results. The next free slot now goes to the company with the fewest runs in progress, so one company's burst no longer holds everyone else's runs back.

​

Improvement. Only Company Admins.
​

Agents that look up travel alerts through the Passolution Global Travel Monitor connector now get a leaner record for each event by default: its id, title, risk level, dates, country and category codes, description and location names. Results are about half the size. If one of your custom agents or skills relies on a field that is no longer included — coordinates, the nationwide flag, the event source, or country and category names — have it request the complete record with sections: ["full"].

​

Improvement.
​

When an agent stops a report on purpose — rather than deliver something incomplete or wrong — the report now shows a Stopped badge and a banner saying so, in the library, in the schedule list and a schedule's run history, and on the report itself. The library's Status filter lists them under Stopped and keeps them out of Ready. A stopped report still opens and exports normally. A scheduled report that stops this way is not emailed to its recipients; the user the schedule runs as gets a notice linking to what the agent produced instead.
​

Learn more in the Reports guide.

​

Improvement. Only Company Admins.
​

The INPUT guardrails now read the text of files attached to a chat message — PDF, plain text, CSV, Markdown and HTML — alongside the message itself, so an instruction or a secret planted inside an upload is judged by the same classifier that judges the typed text. Images and PDFs without a text layer are not screened. Nothing changes for a benign attachment: the file still reaches the agent as before.

​

Fix.
​

Re-opening an agent that follows the company default model left its Vendor blank, so every save failed with "Vendor is required" until you picked one again. The vendor now fills in on its own.

​

Fix.
​

Leaving an optional number input blank on a report's run form showed "Required" and refused to start. It now runs without that value.

​

Fix.
​

Jumping to a section of an agent's configuration could highlight the one above it instead. The same fix applies to the user and company pages.

​

Fix.
​

When a connector refuses an agent because the signed-in identity lacks the required permission, the agent now says so instead of suggesting you try again in a moment. A permission is not something that reappears on its own, so the old wording sent people back to the same question minutes later, and made a configuration problem look like a temporary outage.
​

Refusals that really are temporary, such as hitting a rate limit, are unchanged and can still be retried.

​

Fix.
​

A report could finish generating yet never be saved: the run stayed on "Running" and was later marked failed, because an unrelated background step, such as naming the chat, stalled at the same moment. That no longer happens.