What we shipped

When an AI agent asks for a public page of wagt.ai, it now receives a clean markdown summary, in Arabic or English: the general idea and where everything is, without digging through the page layout. Your browser keeps seeing the site exactly as before.

Wagt reads the comments on your YouTube videos and the channels you follow, picks what is actually worth engaging with, and writes the reply in your voice. Then it hands you the list. Every like, every follow, every comment waits for your tap on that one card. There is no approve all button and there will not be one: the platforms require the person to authorise each action themselves. So Wagt does the finding and the writing, and you do the acting.

Before Wagt can comment on a LinkedIn post it has to read it, and reading it used to ask your permission, because the word post inside the tool name read as publishing. Looking at something cost you the same confirmation as publishing it, which made read then reply unusable. A read is a read now. Publishing still asks, every single time.

Five apps were handing Wagt a search box instead of a list, so it could not see its own abilities and described them from memory instead. That is where confident answers about things it could not actually do came from. Discord, YouTube, Instagram, Outlook and Google Meet now hand it a named, checked list. It answers with what it really has, and it gets there faster.

Deleting your account erased our record of which app permissions were yours, but the permissions themselves stayed alive at the connector. Nothing could reach them afterwards and nobody could find them to cancel them. Every live permission is now revoked before anything is erased.

Ask for a spreadsheet, a deck, a report or a quote and you get something you can open and forward: Excel, PowerPoint, Word or PDF. Arabic is laid out right to left with a real Arabic face, not Latin typesetting with Arabic pushed into it.

Send a clip and ask about what is in it. Wagt watches it rather than reading the file name and guessing. Long or very large clips are still refused rather than quietly half read.

There is now an About page at wagt.ai/about: why we built Wagt, what we hold to about your language and your data, and who it is for. The footer carries our X, LinkedIn and TikTok next to it.

Ask for an invoice's due date and a reminder on that date. If the calendar turns out to be read-only, the old report said only "reconnect your calendar" and the due date it had already found never reached you. You were told your permissions were wrong and asked to start again for work that was already done. Now the proven half comes back with the explanation.

Wagt used to find the invoice, then doubt itself and run the same search again a day wider, then again with no date limit, and open its report by saying it found nothing. Once a set of keywords comes back with results, those keywords do not run again. A search that genuinely found nothing can still be widened.

Days left was counted from a hardcoded seven days instead of the date your card is actually charged, so it drifted by however long the confirmation took to arrive. It now reads the real date.

It said checkout was not open and gave you nowhere to go. True while payments were off, and quietly costing sign-ups once they were on. It now takes you to the plan picker.

Wagt reads a line marked as coming from the team that builds it, and trusts it. The sender, subject and preview of an incoming email were being placed into that same view without escaping, so anyone who could email you could write that line and be believed. The shape is now stripped from anything an outsider wrote.

When Wagt decides a message needs no reply, it uses a tool that says nothing. That tool was being treated as a sensitive action, so it asked for confirmation, failed, and replaced the silence with "I couldn't complete that action" on a turn whose whole purpose was to say nothing.

Checkout now happens inside the dashboard instead of sending you elsewhere. Upgrade, downgrade, cancel, and resume a cancelled subscription before it lapses. A downgrade waits for the end of the period you already paid for.

The screen reported success as soon as the request went out, not when the charge actually settled. If the card was later declined you had already been told you were on the new plan.

Automatic renewal was left to the default instead of being set on every new subscription. The failure mode had no error and no message: the subscription simply ended one day. It is now set explicitly, and rejected outright if the annual price is not configured.

The rules it works from were dense bullets that read once and then get skimmed. They are now plain sentences, one idea per line, with nothing dropped. Folded in: an explicit list of what you can and cannot see, an order to trust when two sources disagree, honesty about a history with gaps, and permission to stay silent instead of filling the space.

Ask Wagt what it does with your data and the answer is now grounded in the published policy rather than written fresh each time.

Moving mail to trash can be done by deleting it or by applying the trash label. The confirmation gate covered the first and not the second, so a bulk clear-out could run through the uncovered route without asking. Both now ask.

A combined security and readiness review before opening payments. Every finding was reproduced by running the real code, not by reading it.

Every dispatched task now carries the conditions that make it finished, and the run spends a step citing the result that proves each one. A missing proof, a malformed one, or a claim with nothing behind it all count as unmet. "I'll summarise what I found so far" is no longer a finished task.

Hitting the ceiling used to end the task there. It now hands the next run what it already proved, the partial report, and what it had surfaced, and continues. Capped at two continuations, and never for scheduled automations or anything waiting on your confirmation. You still see one task, not three.

A playbook for every app in the catalog: the first move, how to phrase a query, the traps, and what result proves the work is done. Every tool name was checked against the live app before the playbook was written, and the ones that do not exist are written in as "do not guess this". Connect an app and its playbook comes with it.

Start linking an app, close the window halfway, and the connection sat as pending with no way out. It is now cleared automatically so you can try again.

Safari returns nothing when a link is opened a particular way, so the tab opened and stayed empty with no way back.

Four allowed tools had been retired by the provider without notice, and one of them was the tidy-my-inbox capability. Asking for it failed in a way that looked like Wagt refusing rather than a tool that was gone.

When a long task continues into a second run, the report from the first is handed over. It is now labelled as material to read rather than as instructions to follow, so text that arrived from outside cannot steer the continuation.

Daily and weekly limits went up by half, to 56 a day and 252 a week. A tester hit the daily wall on nine cents of real cost, which meant the wall was in the wrong place, not that they were using too much.

On phones the browser often blocks the new tab, and the button did nothing at all. It now falls back instead of failing silently, and there is a button for it in the sidebar and the menu.

One tool name that the provider had retired was enough for it to reject the whole session, so no app worked for anyone rather than that one tool failing. The retired names are gone, and a session rejected for a bad name now rebuilds itself instead of staying down.

If the account was not linked yet, the link dropped you into a bare chat with nothing to do. It now only opens once there is something waiting for you.

A test suite of 68 real conversations, in Arabic and English, that a change has to survive. The important ones are run five times, not once, because an agent that gets it right four times out of five is not fixed. Anything that breaks a critical scenario does not ship.

A second agent reads the transcripts and judges whether the reply actually reads like a capable assistant: no leaked machinery, no false confidence, the right language. Getting the task done and sounding wrong is still a failure.

Invited testers get 30 days instead of the standard 7, no card. You were promised a month and the code now says a month.