You can also browse, search and filter all release notes directly in the application → Open release notes in TAI.
Skills: try out or demo itinerary agents on ready-made reference bookings
New Feature.
Want to try out or demo an itinerary agent without a real booking? A set of ready-made reference bookings now ships with TAI: ask Itinerary Checker to run one and it checks it end to end, exactly as it would a live booking — a 15-day family trip through Iceland, Seattle and British Columbia, or a short two-leg European connection. Every response is clearly marked as a sample, and any agent equipped with the new Sample Itineraries skill can use the same set.
Itinerary Checker running the TAIDEMO-VANCOUVER sample itinerary — the reply states up front that it's a sample, not a real booking, then checks it end to end like any live trip.
A generated report carries the same notice inside the document itself, so the disclosure travels with the file.
Company admins can now preview Corporate design live
New Feature. Only Company Admins.
Settings → Corporate design now shows a live preview beside the form — unsaved edits included. Change a colour, font, or radius and watch a sample report page (header, KPI cards, a table, the five severity pills, and inline code) update instantly, so you can tell whether a brand looks right before saving it.
The preview shows the on-screen look; exported PDFs always print on a white background and don't use the Max width setting. Tone of voice shapes the agent's wording, not the page, so it doesn't change the preview.
Ask for a Reports dashboard and still get a complete PDF
New Feature.
Ask an agent for a dashboard rather than a document and you now get one — tabs, expandable sections, sortable tables — instead of a report that only reads top to bottom. The agent picks the shape from what you asked for; there is nothing to configure. What does not change is the PDF: everything behind a tab or a collapsed section is printed open, so the exported and emailed copies still carry the whole report. Scheduled reports keep the shape their agent is set up with, so a weekly report does not look different from one week to the next.
Open a Report on its own page, without the Workbench around it
New Feature.
A report now has its own URL. Open standalone on any report gives you the document full-height with no sidebar and no navigation — the view to use when you are presenting it on a screen, reading it on a small display, or sending the link to a colleague.
Scheduled report emails land here too. The link in a delivery email now opens the report itself rather than the run's details page, which is what you wanted from that click.
Anyone opening the link still signs in to TAI and needs access to that report; this does not share it outside your company.
Reports: interactive reports now explain their own PDF
Improvement.
An interactive report changes as you click through it, so a PDF of it was only ever a snapshot of one state. TAI no longer hands you one:
The report's Export PDF button is gone; Export HTML is the full report.
The chat bubble offers the HTML download alone.
A scheduled email says the report is interactive instead of the generic link note.
A schedule's run history shows, per run, what happened to its PDF.
See Using reports and Scheduling reports.
A schedule's run history now shows why each run shipped no PDF.
Reports: plainer text reports now look like the styled ones
Improvement.
Some reports come back as plain text rather than a designed document. Those now render with the same styling as the rest — your brand's colours, headings, tables and code font, from one shared stylesheet instead of a second, slightly different one.
When you print one, or save it as a PDF from your browser, external links now show their web address after the link text, so a printed copy stays usable.
Knowledge Sync history now shows every run, live
Improvement.
Sync history lists every run — queued, running, done, failed or interrupted — not just the completed ones, each with its start time, duration, trigger and outcome. A no-op re-crawl reads as Index already up to date — no changes, a real one reports the chunks added and removed, and a failure shows its reason. A re-sync shows up the moment you click it, and clicking again while a run is queued says so. Learn more in the Knowledge guide.
Reports now read as a page, not a sheet of paper
Improvement.
Every report used to render as a fixed 880px card on a grey background — the right shape for something you print, the wrong one for something you read on screen. New reports now use the width of your window, while paragraphs stay at a comfortable reading width so long text is no harder to read on a wide monitor.
This applies to reports generated from now on. A report is stored as a finished document, so the ones already in your library keep the appearance they were created with.
Export PDF is unchanged — still A4, same pagination.
Settings → Corporate design → Max width now sets the reading width for text rather than the width of the whole page, so tables and charts can use the full window.
Unexpected errors now say they're ours, and carry a request id
Improvement. Only Company Admins.
An unexpected failure used to surface as a bare "An internal error occurred", which reads like something you did wrong: connector authors were rewriting working code to chase it. It now says the fault is on the TAI platform, that changing your input won't work around it, and quotes the request id to report.
Documentation: printing an article now prints the article, not the whole screen
Improvement.
Hitting print (or Save as PDF) on any documentation page now outputs just the article. The navigation sidebars, top bar, breadcrumbs, and banners are left off the page, and the text uses the full page width.
The article is no longer cut off after the first page — long docs flow across as many pages as they need, and diagrams, images, code blocks, and table rows are kept whole instead of being split at a page break.
Printing from the dark theme now produces readable dark-on-white pages rather than near-invisible light text, and wide code lines wrap instead of being silently cropped at the paper edge.
Advisor now says when your report requirements were the thin part
Improvement. Only Company Admins.
A report's contract has two halves — the agent's prompt and the requirements you typed when you ran it — and the advisor judged both but only ever suggested changing the agent. So a run that failed because "booking failures" was never defined came back proposing a rewrite of a shared system prompt, which is a heavy fix for a one-off report and no fix at all when the agent is a built-in one you cannot edit.
Findings can now point at the run's own requirements, quoting them, when those were the underspecified half.
Where a gap could be closed either way, the advisor names both remedies and the trade-off: changing the prompt is durable and reaches every future run, changing the requirements fixes this run in seconds.
Knowledge: re-syncing a source now only re-indexes what changed
Improvement. Only Company Admins.
Until now a re-sync rebuilt a source's whole index even when almost nothing had changed. It now compares the freshly fetched content against what is already indexed and writes only the pages that are genuinely new, edited or gone, so a routine re-sync of a large source finishes much faster. A sync that is interrupted part-way also picks up where it left off on the next attempt instead of starting over. A sync you start yourself — saving a source, or pressing Re-sync — is now picked up ahead of the scheduled background refreshes, so it no longer waits behind them.
Company admins can now delete Knowledge that is still in use
Improvement. Only Company Admins.
Deleting a knowledge item or collection that something still references used to fail with a message listing what was in the way, leaving you to detach it by hand before you could try again. Now the delete asks instead: a second dialog names what still uses it (the agents using a knowledge item, or the knowledge items, agents and skills attached to a collection), and Delete anyway goes ahead and detaches it from all of them. Skills and connectors (Nezasa staff) work the same way.
Corporate design: a heads-up when a colour pair may be hard to read
Improvement. Only Company Admins.
Saving a corporate design now measures the text pairs a report prints — ink and the softer ink on each of the three surfaces they print on (the page, its alternate surface, and the tint behind table headers and KPI cards), accent on that tint, and the severity text on each of the four severity colours and on the neutral pill fill — against the WCAG AA minimum of 4.5:1 — see what WCAG AA means. The form lists any failing pair and its measured contrast as you edit, so you know which two colours to move — it's a heads-up, not a block, so you can still save. Deriving a palette from your website flags it on arrival too. A dark theme works — it just has to move together: a dark accent tint alongside the dark surfaces, light inks and accent, and light severity fills under a dark severity text.
The default muted text colour has been darkened one step so the shipped theme passes every pair it now checks — previously the small label on a KPI card fell just under the mark. If you have not overridden it, captions, table labels and footers in your reports will read a little stronger than before.
The live warning as you edit: the failing pair and its measured contrast, a link explaining WCAG AA, and Save still enabled.
Chat: attaching an image no longer eats the conversation
Fix.
An attached image was counted at roughly a hundred times what it actually costs the model, so a single photo or screenshot could fill the whole conversation budget and push earlier messages out of it.
Attachments are now counted the way the model charges for them. An image costs the same whether it is 300 KB or 2 MB, and the context meter reads honestly again.
Text attachments still have a ceiling: a CSV, log or JSON file much over 800 KB is refused before the run starts, because that much text genuinely is more than one conversation can hold.
Corporate design: deriving from a website now picks your real brand colour
Fix. Only Company Admins.
Deriving a corporate design from your website could propose a bright secondary colour — one too pale to read as a heading or a link — instead of your actual brand colour. Colours are now chosen by where the page gives them authority (its declared theme colour, then colours its own stylesheet names as the brand) rather than by how often a value happens to appear, and a proposed accent must be readable on a white page before it can be suggested at all.
Font and corner-radius proposals no longer come back empty on sites that define their styling through CSS variables — the underlying value is read instead of the variable reference, and multi-word font names are quoted so they survive.
The font proposal is now the typeface your site actually sets for its text. It no longer picks up the invisible stand-in font a site loads to stop text jumping while the real one downloads, which came back as an unusable name nobody has installed.
The pale accent is now a tint of your brand colour rather than of whatever else the page happened to use, so table headers and KPI cards no longer clash with the headings printed on them.
Corner radius now proposes only a real, card-sized corner. It no longer picks up the fully-rounded value off an avatar or a badge (a
50%or50remrounding), which turned every card and table in your report into a lozenge.Sites that serve their stylesheets from a neighbouring subdomain (for example
assets.yourcompany.com) are now read, where before the derive could come back with nothing at all. Deriving also identifies itself to the site it reads, so sites that turn away unnamed visitors — which failed outright before — now work.If the AI's answer can't be read in full, the roles it did assign are kept rather than discarded, and the review panel tells you that happened.
When deriving can't answer a field it now says so instead of going quiet. If no colour on your site is readable enough to be an accent, or there's no pale tint of it to sit behind text, the review panel names that — previously the field silently kept its old value and the derive looked like it had agreed with it. The same goes for a site linking more stylesheets than can be read in one pass.
Colours and fonts your site defines once and reuses everywhere are now read. Most sites declare a colour as a named token and refer back to it on every rule that paints it; deriving followed only the literal spellings, so it saw the definition and missed all the use that says which colour matters. Those references now count.
The font proposed is the one your body text is set in, not your headline face. A heading-heavy page mentions its display font more often, which used to be enough to win the field and set every paragraph of a report in it.
Sites that are slow to serve their stylesheets now derive consistently. Large sites often sit behind bot protection that answers on a slow path, and the reader gave up too early — so the same site could produce a good proposal one moment and a thin one the next. Stylesheets are now read in parallel and given longer to answer, within the same overall time limit as before.
Muted brand colours are proposed again. A restrained corporate navy or slate is a real brand colour, and the readability check no longer rejects one for being insufficiently bright — while a large muted background is still not mistaken for the brand colour sitting on top of it.
Deriving now shows a progress bar while it works — it reads your site and analyses it, which takes a few seconds — instead of looking idle.
Reports: severity pills and column headers stay whole in the PDF export
Fix.
Exporting a report to PDF could break a short label across two lines mid-word — a MEDIUM severity pill printing as MEDIU / M, a Leg column header as Le / g. The PDF is the copy that gets forwarded to a supplier or filed against a booking, so it is the one that has to read cleanly.
A severity pill and a column header now print as one word, in every report layout and in a report exported earlier — the rule is applied at export, not baked into the stored document.
The same holds for a report written as Markdown rather than as a styled document, where a status value carries no pill of its own.
A table of long values (ids, emails, references) still shrinks to fit the page rather than running off the right edge.
Prompt Advisor runs name the advised agent even once it is deleted or suspended
Fix. Only Company Admins.
The Prompt Advisor run list and a run's detail page name the advised agent even after it was deleted or suspended, marking its status instead of showing a raw identifier. A deleted agent no longer offers to apply or draft a proposal, while the before/after comparison — which only reads recorded conversations — stays available.
A proposal whose current prompt is not visible to your account now says so, rather than previewing the rewrite as if the agent had no prompt at all.
Connectors that return a JSON list no longer fail
Fix. Only Company Admins.
A connector whose endpoint answers with a top-level JSON array (a list of public holidays, countries, bookings) failed every call with a generic internal error, in the Connector Lab and in live agent runs alike. The response contract accepted only a JSON object; it now accepts whatever JSON the upstream sends. Nothing to re-author: existing connectors start working on their next call.
Connector Builder's review panel now updates without a reload
Fix. Only Company Admins.
The builder agent works for minutes at a time, so the tab watching it is often not the focused one. The verification report and credential panel stopped polling whenever the tab went to the background and stayed frozen on "Waiting for the assistant" until the page was reloaded by hand. They now keep polling in the background and pick the agent's submission up on their own, and the Connector Lab's invocation trace refreshes as soon as you come back to its tab.
Long report runs no longer stop part-way through
Fix.
A report run that went quiet for more than two minutes — a slow data call, a long stretch of thinking — lost its connection and was recorded as interrupted. The connection now stays open for as long as the run takes.
Report runs now end when the agent finishes
Fix.
A run could keep going after the report was already written — re-planning and re-rendering it for several more minutes, at several times the credits, and sometimes ending as “the report grew too large to finish” for a short request. Runs now stop when the agent says it is done.
Advisor: a proposed prompt can no longer change what no finding raised
Fix. Only Company Admins.
A proposed prompt rewrite could quietly edit sections none of its findings mentioned — condensing guidance, merging and renaming sections, and in one case dropping an instruction telling the agent not to share sensitive system information — while its own rationale said it had left everything else alone. The diff preview showed it, but only to someone who read the whole diff instead of the summary that told them not to bother.
A proposed prompt is now an edit, not a rewrite: anything no finding raised survives word for word, including section order, names and formatting.
The rationale must name every section it did change, including any it merged, renamed or moved — so the diff can be checked against it rather than taken on trust.
Advisor: a report run with no output is no longer blamed on your prompt
Fix. Only Company Admins.
Analyzing a report run that failed before the agent produced anything — no reasoning, no progress, no report — could come back with a confident critique of the system prompt and a proposed rewrite. There was nothing to judge that prompt against: the run emitted no output, so any prompt would have produced the same empty result, and the critique was unfalsifiable. Worse, the same kind of failure could be blamed on the prompt in one run and on the platform in the next, so a single verdict told you nothing about whether to trust it.
A run that produced no output is now reported against the error it actually recorded, with no prompt rewrite proposed.
A run you stopped — cancelled, or the tab closed while it was still starting — is reported as exactly that, and asks you to re-run it. It is no longer dressed up as a fault worth reporting.
A scheduled run that stopped the same way is reported the other way round: nobody had a tab open on it, so it is the platform losing your run and it says so.
An on-demand run that stopped that way could be either — you navigated away, or it ran in the background and hit its time limit. The advisor now says so, instead of picking one and sounding sure.
A run that genuinely broke leads with the platform error, so you meet the cause before anything else.
Blaming the prompt now requires evidence that the agent acted and acted wrongly.
Gaps the advisor can see in the prompt itself are still reported — a platform stop hides the cause of this run, not a contract that was thin to begin with.
Agents no longer read a settled TripBuilder commission as already paid
Fix.
"Settled" says a commission was recorded on the order, not that it has been paid out. Agents were reading it as paid in full with nothing outstanding; they no longer do.
Chats: Connector Lab and Builder sessions no longer clutter your conversation list
Fix. Only Company Admins.
The Connector Lab and the Connector Builder run on internal system agents, and the chats those pages create were showing up in Recent Conversations, the Chats sidebar and the Starred tile as if you had started them. They are now hidden from every session list.
For customers those rows were dead ends: opening one failed, because the system agent behind it is deliberately not reachable from the chat terminal, and there was no way to remove the row from the list either.
Nothing is lost. The Lab and the Builder still resume your last conversation from their own pages, exactly as before.