You can also browse, search and filter all release notes directly in the application → Open release notes in TAI.
Batch agents can process far more bookings before hitting turn limits
Improvement.
Custom agents that loop over many bookings (passenger manifests, flight exports) used to stall around 40 itineraries: unless every single call asked for specific response sections, each fetch pulled the full booking bundle (~44 KB), rapidly exhausting the turn's processing budget.
Lean by default — an itinerary-context fetch without a
sectionsselection now returns the lean default (trip_facts+components_overview, roughly 3-6 KB) instead of the full bundle, a ~10-20x cut per booking. Explicitsectionsselections work exactly as before, andsections: ["full"]fetches the entire bundle when you genuinely need it.Batch guidance built in — the Cockpit API skill now teaches batch jobs to size
sectionsto the columns they extract and to re-write their output file after every batch, so even a run that hits a processing limit delivers the rows completed so far.
TripBuilder agents can pull booking-agency contact details in one call
Improvement.
The TripBuilder agents — Ops & BI, Itinerary Checker, Trip Guard, and Passolution — can now look up travel-agency contact details (phone, email, postal address, website) through a single call to the Cockpit API, instead of opening one itinerary at a time to harvest them.
Crisis reports that need to reach the booking agency — the trigger for this change — assemble in seconds rather than timing out on large affected-booking lists.
The postal address, which the itinerary bundle does not carry, is now reachable for the first time.
The skill also tells an agent that already holds an itinerary to read the agency's phone and email straight from that bundle — no extra call — so the new endpoint is only touched when address or a stand-alone agency list is genuinely needed.
Answers no longer cut off silently after data lookups finish
Fix.
In rare cases an agent could finish all its data lookups and then stop right before writing the final answer — mid-sentence or with nothing at all — while the chat showed a normal, complete turn. The platform now catches more of these stalls and automatically prompts the agent to finish its answer; if a turn still ends incomplete, it is clearly marked as incomplete instead of looking like a finished reply.
