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

​

Major New Feature. Only Company Admins.
​

The best evidence of whether an agent works is what it actually did. The new Advisor evaluates recorded conversations against a quality rubric — task completion, dead-ends, tool and data-source usage, tone, prompt complexity — and returns findings with links to the exact messages, plus one-click-applicable proposals: a rewritten system prompt (shown as a diff before anything changes), skill attachments, or a drafted custom skill. It can also analyze a report run's contract and failure causes, and predict — turn by turn — how a proposed prompt would have changed a recorded conversation before you adopt it.
​

Runs execute in the background: start one from an agent's Advisor card or a report run, navigate away freely, and find the results under the new Advisor section. Company admins only; runs are capped per day.
​

The Advisor card on an agent's detail page: pick recorded conversations, add an optional steer, and start a run.
​

A finished run: summary, findings with severity badges and links to the exact messages, and the proposals you can apply.

​

Improvement. Only Company Admins.
​

The four settings that decide what AI does with your data now sit together at the top of Settings, under AI & Data: AI model defaults, Web access, Data access by Nezasa, and Report delivery. Model region and the default capability tier used to be a card inside Company, where nothing in the navigation admitted they existed — they are now their own page. Data access is renamed Data access by Nezasa, so the label says whose access it grants. The remaining sections are Users & Access and Workspace.
​

Settings now opens on AI & Data, with Users & Access and Workspace below it.

​

Improvement.
​

Switch Release Notes between Detailed and Compact at the top of the page. Compact collapses each release into New / Improved / Fix sections with one line per change — the same shape as the summary we post when a release goes live. Click any line to open that note in full, screenshots included, without leaving the overview.
​

The choice rides in the URL (?view=compact), so a link you share opens on the view you were looking at.
​

Compact view: one line per change, grouped by New / Improved / Fix, with the Detailed/Compact switch at the top of the page.

​

Improvement.
​

The agent advisor grew beyond system prompts. It now reviews an agent's configuration too — model fit, attachment hygiene, welcome copy — alongside the existing prompt operations. And every proposed rewrite is shown as a diff against your current text, with severity badges on findings, so you see exactly what would change before applying anything.

​

Improvement.
​

Asking the Traveller Locator for detail checks every matching booking in parallel, and each one it opens takes up room — so the drill used to run out of room part-way down a long list and ask you to reply continue.
​

It now asks TripBuilder for just the trip outline and the flight details it actually prints, instead of the entire booking record. On a 19-service booking that cut what each one costs by about 83% — from roughly 81 KB to 14 KB — so a pass gets through more of the list before it stops. The table itself is unchanged: same passengers, same contact details, same hotel, same first and last flight with PNR.

​

Improvement.
​

Two improvements to how agents work under the hood.

​

Fix.
​

The companion Skill's description and content already followed a connector rename; its title now does too, instead of keeping the old name.

​

Fix. Only Company Admins.
​

​

Fix.
​

An agent that reached for two or more connectors in the same step could end the run with a provider error instead of an answer, and a scheduled report doing so delivered a failure notice rather than the report. The step's tool results were being sent in an order the model provider rejects; they are now kept together at the front of the step, so the run completes normally.

​

Fix.
​

Editing a report schedule — even just its name — used to silently flatten a JSON-shaped input into text. It now keeps the value it started as. JSON inputs also get a roomier, monospaced field that flags invalid JSON as you submit, rather than accepting it and failing later.

​

Fix.
​

The Unknown connector row's ids were reachable only by hovering — nothing on a touch screen, and unannounced by a screen reader. They now print under the row.

​

Fix.
​

An agent that saved a file outside the folder a run collects downloads from left no trace at all — no download, no explanation, under a reply that said it was saved. It is now named with the reason it was not delivered.

​

Fix.
​

The header bell used to open a preview of the newest dozen notes, but opening it marked every unread note read — including the ones the preview never showed. The bell now takes you straight to Release Notes, and the count clears once you're on the page with the whole list in front of you.

​

Fix.
​

A flight can be cancelled, on request, or never confirmed by the supplier while the booking itself stays confirmed. The locator used to print those in the first- and last-flight columns with their booking reference, which reads as a held seat. They now say Flight cancelled or carry an (unconfirmed) marker instead, and each one raises a flag in the summary.

​

Fix.
​

Leaving an agent's Model on Follow company default used to fail on save. It now works on create and edit, and the agent tracks whichever default your company is on.

​

Fix.
​

The title stamps the run's UTC start, while Created and Finished show your local time. It now says which is which. Existing report titles are unchanged.