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

​

New Feature. Only Company Admins.
​

Settings → Rate limits shows the requests-per-minute allocation Nezasa has granted your company, and lets you tighten it: a lower company-wide cap, plus the default allowance any one user or API key gets — the same allowance that covers chat messages, report runs, knowledge queries and agent invocations. Every REST-connected agent is listed on the same screen, so you can give one its own limit, change it, or put it back on the shared default. See Rate limits.

​

Major New Feature. Only Company Admins.
​

The Billing Center's new Budgets page lets you cap billable credit consumption over a day, week or month — up to three budgets at once, one per period. Set alert thresholds and TAI emails your chosen admins the first time each one is crossed, once per threshold per period.
​

Turn on a budget's hard stop and it also refuses new chat turns, report runs and agent invocations once the limit is reached — a banner explains why and when it lifts, and runs already in progress are never cut off. Start with alerts only, and turn the hard stop on once you're satisfied with how they behave.
​

See the Credit budgets guide for what counts towards a budget, why usage can still overshoot it, and how resets work.
​

Three budgets, three states: on track, warning, and exceeded and blocked.
​

Set a budget, its alert thresholds, recipients, and an optional hard stop.

​

New Feature.
​

If a company admin has set a credit budget with a hard stop, reaching it pauses new chat turns, report runs and agent invocations until the budget's period resets. Instead of a request that just fails, you now get a banner saying usage is paused and when it lifts — in chat, on the Reports page, and anywhere else a run would have started.
​

Anything already running finishes normally. The pause clears by itself at the reset, or as soon as a company admin raises or turns off the budget.

​

New Feature.
​

A new guide under Integrations walks through calling an agent's REST API channel end-to-end: authenticating with an API key, discovering an agent's versioned contract, invoking it, reading the response envelope, and handling every error code and rate limit. The channel itself is still managed by Nezasa today and will open to company users in a future release — this guide lets your integration team get ready ahead of that.
​

See the full guide: Agent REST API

​

New Feature.
​

An agent's Reports channel now has a Layout setting beside the output format. Leave it automatic and each run decides, or pick Document for something read top to bottom and Interactive page for something to scan and explore — tabs, filters, sortable tables. The choice is made once per agent and every report that agent writes follows it, in chat and on a schedule alike.
​

Both layouts export and email as a PDF. An interactive report prints with its tabs and panels opened out, so the PDF carries the whole report rather than the one view you were looking at.
​

Layout applies to HTML reports; a Markdown channel is always a document.
​

See Using agents.
​

The new Layout setting on an agent's Report channel.

​

Improvement.
​

​

Improvement.
​

Booking Completeness used to recognise only open, on-request and pending services. It now reads every booking status a service can carry, so a cancelled service on a booked trip is reported as a HIGH finding that names it — rising to CRITICAL within a week of departure, when there is no longer time to re-source it — unless another booked service already covers the same leg or nights, in which case the check stays quiet, because someone already re-protected the trip. Failed and non-bookable services are now named as definitive outcomes rather than merely unconfirmed ones, and an unknown status is flagged as a possible ghost booking. There is also a new sample itinerary, TAIDEMO-CANCELLED-LEG, to try it on.

​

Improvement.
​

Each list remembers the maturity, status, ownership, type, and sort selection you last used — independently per list — and applies it automatically next time, instead of always starting from "All". The Chats agent picker now works the same way, remembering your tier and maturity filter. Change or clear a filter and the new selection is what comes back next time. The selection is scoped to you, not shared with your company, and syncs across your own devices. Search terms are never remembered — only standing filter choices.

​

Improvement. Only Company Admins.
​

The Advisor runs list, the run detail page and the session pickers behind them now load noticeably quicker, especially on agents with a long history: each page fetches only what it shows instead of the full result of every run it lists.
​

The before/after preview on a Replace prompt proposal also renders faster on long prompts, so opening a proposal no longer pauses the page while the comparison is worked out.

​

Improvement.
​

Two more starter prompts on the TripBuilder Ops & BI Agent's chat start screen. One asks for a portfolio snapshot — trends, top destinations, revenue against volumes, and service mix, with charts. The other asks what needs acting on: unpaid trips departing soon, action-required items, and payment risk across confirmed bookings. Learn more.

​

Improvement. Only Company Admins.
​

An agent's page is now split into a Configuration tab and an Advisor tab. The Advisor used to sit below the full-width Save bar, which reads as the end of the page, so reaching it meant scrolling past every configuration section first.
​

The Advisor tab also gained structure: past evaluations and starting a new evaluation are now separate sections, and an agent nobody has evaluated yet says so instead of showing nothing. Built-in agents get the same two tabs, and the open tab is part of the page address, so you can link straight to it.
​

The Advisor tab on a built-in agent, with past evaluations and a new evaluation as separate sections.

​

Fix. Only Company Admins.
​

Suspending a user who was an alert recipient on a credit budget made that budget impossible to save — every edit came back with "recipient … is not an active user of this company", even when you were only changing the credit amount. The same happened to a recipient who simply lost the Company Admin role. Editing a budget after an ordinary offboarding now works.

​

Fix.
​

When a report run can't produce a usable report, it now ends as Incomplete with a plain-language reason, instead of publishing whatever text the run happened to leave behind. A provider timeout or outage shows a short, readable message rather than a technical error, and a failed run links straight back to its own run page. And an Excel or Word export that would be too large to build is refused with the limit named, so you can narrow the request and re-run.

​

Fix. Only Company Admins.
​

Crossing a budget threshold could occasionally email the same alert twice. An alert already being delivered is now no longer picked up and sent a second time.

​

Fix.
​

​

Fix. Only Company Admins.
​

A conversation you came back to over several days had all of its credits recorded against the day it started, and usage added to it after a month was finalised could go uncounted altogether. Every day's usage is now priced on the day it happened, so daily and monthly figures add up whichever days a conversation spans. The CSV export follows the same rule: a conversation that ran across several days now has one row per day. See the Billing Center guide.

​

Fix. Only Company Admins.
​

The prompt advisor's Review configuration check said it judged an agent's channels, but the channel list it received was always empty — so its consistency check could never fire. It now receives the channels the draft actually has switched on (REST API, Streaming API, MCP, Report), and will flag a system prompt that contradicts them: an agent that tells users it only answers in chat while the Report channel is enabled, for example.

​

Major Fix. Only Company Admins.
​

REST agent invocations and Prompt Advisor runs now count toward usage and credit budgets, matching what the Credit budgets guide has always described. Rates are unchanged, but this traffic previously counted as zero.
​

Each call is billed as its own session. One REST invocation, or one Advisor run, is one session — each with the same per-session minimum a chat has. Credits are therefore charged per call rather than pooled across them, so a thousand small machine-to-machine calls cost a thousand minimums rather than the handful of credits their tokens alone would suggest. High-volume API integrations will see the largest change.
​

If you use the REST API or the Prompt Advisor, expect your Credits to rise and check that any hard-stop budget still has room.

​

Fix.
​

A long Chats turn that consulted several sources could end with nothing but the "ran out of steps" note — everything it had already read was thrown away. Such a turn now answers with what it actually gathered first, naming each source it reached, and leaves what is still missing to a closing line.

​

Fix.
​

A report set to Interactive page could occasionally arrive as a plain document. It now keeps the layout the agent is set to.

​

Fix.
​

​

Fix. Only Company Admins.
​

Hiding a built-in agent's prompt excerpts also dropped quotes from your own conversations and report runs. Advisor runs from now on keep those quotes; the built-in prompt stays hidden. Runs from before this change are unchanged — nothing in a stored result tells the two kinds of excerpt apart, so those keep hiding both.

​

Fix.
​

Tool-using agents at High reasoning effort could still spill follow-up reasoning into the visible answer after running a tool. Reasoning now round-trips as reasoning, so the Show thinking panel holds all of it at every effort level. Built-in agents temporarily pinned to Medium are back on the model's default; agents with an explicit level keep it.

​

Fix.
​

In those sections, a detail URL that isn't a real item — a typo, or an old bookmark — now returns you to where you pick one, not an error.

​

Fix. Only Company Admins.
​

A refused save used to lead with the schema path the backend complains in. It now names the field by the label shown above it, and highlights that control — expanding the panel it sits in.