Start

Novità

Cosa cambia in Anomalia

Ogni aggiunta, miglioramento e correzione al prodotto — dal più recente al più vecchio.

August 26, 2026

Your agent has a real desktop

  • Reopening a chat while an answer is still being written no longer shows scrambled text: the reply picks up exactly where it was.
  • When you watch your agent’s computer, you now see it browse: pages open on the screen instead of somewhere you cannot see.
  • Work an agent hands to a helper now happens on that agent’s own computer, the one the panel shows you.
  • Self-hosting now works end to end from an empty database: bring the stack up, run one command, and you can sign in — schema, background jobs and login all verified on a clean install.
  • The model you pick for a chat now sticks to that chat, on every device — reopen it tomorrow on your phone and it answers on the same model.
  • Each of your custom agents can now be set to a specific model, and it uses that model on its scheduled runs too.
  • Motion videos asked for in chat are now built by the same engine as the Motion video studio, so they get the full craft — real transitions, easing, reference structure — instead of a first draft.
  • Chats in the sidebar now show a steady green dot on every agent that is still working, right when you open the app — no need to open each conversation.
  • Agents can now see what their teammates have been working on and which teammate conversations are waiting on them, so handovers happen instead of duplicated work.
  • Agents can no longer overwrite work they have not seen: editing a post, document, graphic or motion source requires reading it first, and if it changed in the meantime the agent is asked to re-read and try again.
  • Pricing cards no longer show a leftover “credits included” line under the price. Free now shows the real credit amount instead of a broken placeholder.
  • Your agent’s computer is now a full Ubuntu desktop — panel, app menu, dock, file manager, browser — instead of a blank screen.
  • Click Take control in the agent panel and the desktop opens full screen, with mouse and keyboard on the same windows the agent is using.
  • Copy and paste now works both ways between that desktop and your own device, from the buttons under the screen.
  • On a phone, a Keyboard button brings up your keyboard and types into the desktop.
  • The desktop no longer shuts down under you after a few minutes: it stays up while you are on the page.
  • Signing in to a site by hand now stays on that machine, so the agent can keep working there without ever seeing your password.
  • Your agent now works on one machine instead of three, so what you see in the panel is the machine it is actually using.
  • The desktop now starts in seconds instead of minutes, and turns on reliably — it used to fail to install altogether.
  • Asking an agent to render a motion video works again — renders were failing on a missing piece of the rendering machine, and the error came back as a closed VM.
August 26, 2026

Optional repo skills for agents

  • Agent runs can now load the repository’s markdown skill packs on demand, opt-in via the HARNESS_SKILLS setting (off by default).
August 25, 2026

Agents can delegate again

  • Specialist agents in the new engine can now hand sub-tasks to helper agents (research, execution, verification) — the same delegation that already worked in standard chat, with no helpers spawning further helpers.
August 25, 2026

Motion, edited in place

  • Motion compositions can now be edited with targeted search-and-replace instead of full rewrites.
August 25, 2026

Agent faces move again

  • The sidebar now has a proper edge, and its top row lines up with the page header above the content.
  • Reload the page mid-answer and the work an agent has already done is there waiting, instead of a thread that looks like it was wiped.
  • Switching to another tab no longer freezes a running answer: come back and it has kept going.
  • The status line and its timer now sit at the bottom of a running answer, under the tool steps too, with the agent face beside them for the whole turn.
  • Tool steps in a running answer now keep what they were called with and what they returned, so opening one tells you something.
  • Stop now works after a page reload: it used to look like a button and do nothing, leaving the turn running and your next message stuck in the queue.
  • "Send now" on a queued message works after a reload too, for the same reason.
  • A reply no longer shifts around when it finishes: the agent face, the spacing and the action row stay where they were, so the chat does not jump.
  • The New messages line now only shows up for replies you actually missed — not for one that just came in while you were reading, and not after a refresh.
  • Agent avatars now morph between expressions instead of switching instantly.
August 25, 2026

Self-hosting, for real

  • Self-hosted instances can now run a real production build — start the app from Docker Compose or a Node server instead of the dev server.
  • A built-in scheduler container now runs all background jobs on self-hosted instances, no external cron setup needed.
  • Instances without social publishing connected now skip posting with a clear reason instead of failing.
  • Support email, sender domain, app URL and crawler identity are now configurable through environment variables.
August 25, 2026

Smoother start for every workspace

  • Creating a workspace no longer fails when your brand name uses a non-Latin alphabet — the web address is picked automatically if it is already taken.
August 25, 2026

One engine under the hood

  • Specialist agents now run on a single, shared engine — same behaviour in chat, far less machinery to keep alive.
  • If the AI provider is not configured, an agent turn now fails fast with a clear error instead of silently falling back.
August 24, 2026

Nothing lost on a refresh

  • A correction sent while an agent is working now changes the answer it is still writing, not the next one.
  • Write to a specialist agent while it is working and it now hears you within the same turn, instead of holding your message until the turn is over.
  • Reload the page while an agent is working and you now get a way to stop that turn, instead of a line of text counting the minutes.
  • /goal now works with specialist agents: they break the goal down, tick each criterion off as it becomes true, and pick the work back up on their own until it is done.
  • A specialist that says a job is finished is still checked against what it actually did — having a goal open no longer switches that check off.
  • Ask and Plan mode now hold on specialist agents too: in Ask a specialist can only read, and can no longer create, approve or schedule anything.
  • Specialists can now hand work to each other: the agent who finishes a video can ask the web specialist for the article that goes with it, without you relaying it.
  • Ask an agent to change "the motion video from the 4th" and it now knows which one you mean — the list of videos carries their dates, and where to open the code behind each one.
  • A tool that keeps failing now disappears from the agent’s toolbox for a moment, with a note saying why — instead of the agent retrying it and burning the turn.
  • Specialist agents now end their turn the moment they answer you or ask you something, instead of sometimes talking past their own question.
  • A message sent while the assistant is still working is now heard exactly once — it no longer repeats itself into the rest of the turn.
  • Motion videos built from a studied reference no longer re-attach the reference frames at every step of the work.
  • A message sent while the agent is still working now counts exactly once: it no longer risks being applied twice in the same turn.
  • A specialist agent whose turn dies now frees the chat in seconds instead of minutes — no more waiting on a "working" bar with nothing behind it.
  • A long background turn no longer shows a false error banner while it is still working.
  • The Analyst specialist can now actually do its job: read your strategy, posts, leads and history — and run the analytics review — instead of having no tools at all.
  • The Content Creator can now edit a post you already have: change the caption, move it to another time, or add a platform, without recreating it from scratch.
  • The Motion specialist can now add a voice-over and a music bed to a video it makes, instead of only ever producing a silent one.
  • Specialist agents no longer waste a step opening files that were never there: every document their briefing points them to now exists.
  • Specialist agents now search your uploaded documents and notes by meaning, not just by exact words — so an answer that only exists in a file you uploaded is found and quoted.
  • Specialist agents can now read your strategy, editorial plan and GTM roadmap as one document, instead of piecing it together table by table.
  • Stop now really stops a specialist mid-task: the work halts, nothing new is charged, and no answer shows up afterwards.
  • Whatever a specialist had already done before Stop or an interruption stays in the chat, tool steps included, so a resumed task carries on instead of starting over.
  • A specialist that runs commands on its machine no longer reaches the open internet, so nothing about your brand can leave that way.
  • An agent whose action fails twice in a row the same way now changes approach or says what is missing, instead of retrying the same thing over and over.
  • Work interrupted mid-task is no longer redone from scratch: the assistant now remembers the interrupted step and checks what already happened before spending credits or posting again.
  • A reply can no longer show up twice in chat when a turn is recovered after a crash or timeout.
  • A specialist turn that was already declared dead now also stops working in the background, instead of quietly burning through resources.
  • Deleting a queued message now really deletes it — before, it silently stayed and could still run later.
  • "Send now" on a queued message sends it once and removes it from the queue, instead of duplicating it.
  • The chat no longer stays stuck busy after a turn has finished: new messages send right away instead of piling up in the queue.
  • An agent that catches itself about to repeat the same answer without doing new work now corrects course or asks what is missing, instead of looping forever.
  • Asking for a post on a platform you have not connected now gets an immediate answer instead of a made-to-order post nobody can publish — the assistant tells you to connect it first, and can still prepare a draft if you ask.
  • Approving a post with no connected account for its platform now says plainly that it is approved but will not go out yet, instead of reading like it was scheduled.
  • A long turn that runs out of steps now ends with one answer instead of two messages in a row.
  • Specialists no longer waste a step opening brand files that were never there — they go straight to the tool that has the data.
  • An article the assistant actually wrote is no longer mistaken for an empty promise and redone from scratch.
  • The Analyst now sees everything your brand has actually published — posts from your connected accounts count too, so it no longer reports "nothing published" when there is.
  • Asking an agent about performance no longer wastes a step on a dead end: the published history is a file it can open directly, with the real numbers.
  • A specialist agent that runs out of time on a long task now picks up where it left off automatically, instead of stopping halfway and waiting for you to ask again.
  • Ask any specialist who else is on the team and you get a straight answer: every colleague, what each one does, who they usually hand work to, and how they write to each other.
  • Every specialist can also tell you what it will never do — pay for anything, log in as you, or delete another specialist — instead of quietly working around it.
  • Ask an agent to check with two colleagues at once and it now writes to both in one go — but only when you ask: agents no longer decide on their own to message the whole team.
  • The assistant no longer repeats itself while working — its working notes now collapse under "Thought for a moment" instead of showing up as separate messages.
  • An honest "nothing to report" no longer gets mistaken for a finished job and re-triggered — and a real unfinished promise is now caught reliably, in Italian and English.
  • A long reply now keeps running on our side even if you close the tab or lose your connection halfway — it used to be cut off and never saved.
  • Work an agent had already written before being interrupted now reliably lands in the chat, instead of the turn simply disappearing.
  • A follow-up you queued while an agent is working now starts as soon as that agent finishes, instead of waiting for the next sweep.
  • Sending images while a reply is still in progress no longer sends them silently without the images: the message stays in the box so you can send it when the chat is free.
  • The live progress of an ongoing agent stays visible after you queue a follow-up, instead of being replaced by an empty spinner.
  • Answering an agent’s question hours later no longer risks the reply being cut short.
  • A queued message is no longer posted twice when it arrives just as another reply starts.
  • Open chats are lighter on battery and data: the page stops checking for updates when it is in the background.
  • Chats with a specialist now appear in your usage and count against your credits — those turns were running off the books.
  • Sending the same message again while the agent is still working no longer starts two competing replies — the new message waits its turn.
  • Refresh the page while an agent is working and the live progress now picks right back up, instead of freezing on a spinner that never ends.
  • If your connection drops mid-reply, the reply keeps going on our side and the page catches back up on its own — no more error bar over work that was still happening.
  • A reply stuck on a dead connection now frees itself after a moment instead of staying frozen until you reload.
  • A message that fails to reach us — say your wifi drops as you hit send — is no longer shown as sent: the error stays on screen with a working Retry button until you act on it.
  • A follow-up queued behind a busy reply that never reaches us now goes back into the box and says so, instead of vanishing.
  • Opening a chat with a pre-filled message now survives a refresh at the wrong moment — the message is delivered once, not lost and not duplicated.
  • Sending a message while an agent is still working after a page reload now queues it instead of starting a second, competing reply.
  • Text you typed but had not sent yet now survives a page refresh — it is right there in the box when the page comes back.
August 23, 2026

Agents can find what they were given

  • Reload a page mid-turn and the reply keeps streaming in — it used to just say "still working" until you refreshed again.
  • The Content Creator, UGC Specialist and Web Specialist can now actually create posts, graphics, AI photos, videos and blog articles — not just analyze what already exists.
  • The agent panel now shows a live view of the agent’s screen, updating as it works.
  • Agents can now open a real desktop — taking screenshots and clicking, typing or scrolling to get things done.
  • The Motion specialist now renders the actual video in chat, with a real preview link, instead of just describing what it would look like.
  • When an agent thinks more than once while replying, each stretch of thinking now shows up where it actually happened — instead of every bit of thinking glued into one row above the whole answer.
  • Thinking is now a quiet, clickable row instead of appearing in full above the answer.
  • Replies now say what exists, with its id and where it sits, then stop — no preamble about what the agent is about to do, no walkthrough of how it did it, and no closing “shall I proceed?”. Anything that failed is still explained in full.
  • Agents can now read your data straight from the database when no existing tool covers the question — with your own permissions and read-only, so they see exactly what you would see in the app and nothing belonging to anyone else.
  • A single reply now stops once it has burned through its token budget, and says so, instead of running up an unbounded bill — everything it finished is saved, and you pick which part to carry on.
  • Agents now carry about a fifth less standing instruction in every step, so a conversation stays useful for longer before it has to be trimmed.
  • Your agents no longer carry your whole brand Studio — voice, palette, products, people, competitors — into every step of every reply; they look it up when the job actually needs it, and the general assistant, the makers and the specialists all read the same one.
  • The general assistant now reads the video-making guide before it writes a video, the way the Motion specialist already did, instead of carrying the whole manual everywhere.
  • The playbook for pitching an idea that is not the safe default is now something an agent opens before proposing angles, campaigns or scripts — your brand’s saved ideas still sit in front of it at all times.
  • Long video renders no longer get cut off partway through on brands that had made one before.
  • The check that stops a voice-over from being cut off mid-sentence now measures the video that actually comes out — on most compositions it had been measuring a six-second one however long the real video was, so a line could get clipped with nothing noticing.
  • Agents can now look up a file by its name across everything they hold, instead of opening folders one at a time.
  • Searching now takes about a second instead of twenty.
  • A search now brings back results from every file that contains them, instead of a dozen lines from the first one.
  • A search now says how many files it actually looked in, and names the ones it could not open — it no longer claims to have read them all.
  • Searching somewhere an agent does not cover now says exactly that, instead of an error the agent reads as “that does not exist”.
  • Searching now finds an Italian word however the accent was typed, and finds it without the accent too.
  • Agents can read a long file a page at a time, and are told where they stopped and how to carry on.
  • A long list can now be filtered and its limit raised, instead of showing the first sixty with no way to reach the rest.
  • Skimming the video-making guide no longer counts as having read it: an agent still has to read it in full before it writes a video.
  • When an agent counts something for you, a metric that simply is not collected for your account now reads as missing instead of quietly counting as zero — so a report can no longer show a median of 0 as if it were a real number.
  • Your posts and your strategy are now actually on hand when an agent runs the numbers: both had been failing to load, and the agent was told the data did not exist.
  • Agents no longer double-count a post that was picked up twice from two different sources.
  • Time an agent spends on its own machine is now billed to your account like everything else, including runs that fail — the machine ran either way.
  • Specialists get their brand context back: the Content Creator can see your media library, editorial plan and scheduled posts again, and the Analyst can see your strategy, roadmap, competitors and leads — all of it had silently stopped reaching them two days ago.
  • Agents no longer promise you something they cannot do — every tool named in their instructions is now one they actually hold.
  • The Web Specialist can now show you the covers and images it makes, instead of pasting a link you cannot watch — and a finished video is handed over the same way.
  • Any specialist can now write a large request out as a plan document, which until now only the general assistant could do even though all of them were told to.
  • A file you attach is now saveable into your brand knowledge whichever specialist you are talking to, not only the Content Creator.
  • The Analyst can correct the competitor list it works from, and the Content Creator can rename, refile or delete a document and manage the three style references that steer every generated image.
  • Specialists now know they have their own machine for anything that has to be counted or computed, instead of only being told to ask a helper for one.
  • Asking an agent to score a finished video in chat now gets you its own read of the clip instead of a scoring step that kept failing — the stored score on a post is still shown, and the media reviewer page still writes new ones.
  • Your specialists can now browse and search every guide and reference they have been given, not just the handful named in their index — the motion agent went from seeing one file to all 153, including the full animation documentation it builds videos from.
  • Listing a folder that exists no longer comes back empty, and searching inside one no longer comes back as an error.
  • Opening a folder now shows what is directly inside it — each subfolder with how many files it holds — instead of dumping seventy-four files at once, and the agent can ask for the whole subtree when it actually wants it.
  • Live SEO and keyword research now belongs to the Web Specialist alone — the other specialists ask it for the numbers instead of carrying the whole research kit into every step.
August 22, 2026

Your agents, working together

  • Type /clear in a chat to give the agent a clean slate — it starts fresh from that point, while every message you have exchanged stays in the conversation.
  • When a goal closes on its own, the note now says the system closed it — not the agent — and how many steps were ticked off from the reply text rather than by a tool that ran.
  • Closing a goal now asks for what exists that did not exist before, with ids, or why it cannot be done — repeating the checklist back no longer counts.
  • Ticking off a goal step is now refused when nothing backs it: the agent is told which step, what is missing, and that it can either do the work or declare the step impossible with a reason — instead of quietly marking it done.
  • A step can no longer be ticked off in the same turn where the tool doing the work was refused — it reopens instead, saying which tool did not go through.
  • While a goal is open, a turn that only read things now says so, however it describes itself — “I created it” and “I patched it” cannot stand on a turn that wrote nothing.
  • Saving a video composition no longer reports success with a link to the gallery before the video is rendered — it says plainly that there is nothing to watch yet and what the next step is.
  • When a step asks the agent to look and call it again, it is no longer reported as a failure — so the agent retries instead of giving up and calling the job done.
  • A reply now names any action that did not go through, so “it is ready” and “it was refused” can no longer sit in the same message unnoticed.
  • The actions row now says how many did not complete, instead of only how many were taken.
  • A goal step now closes on what happened, not on the agent saying it happened: a tool that failed no longer counts as work, a turn that only read things cannot tick anything off, and a step that names the tool it needs stays open until that tool has actually run.
  • An agent can no longer hand you a file it did not make — when a reply runs no tool at all, it says so, and any link in it is an older file.
  • When background work fails to start, the chat now tells you nothing is running instead of leaving you waiting.
  • A screenshot of an interface can no longer sit frozen behind a video with things floating on top of it — the check is on by default, and it asks for the product to be rebuilt and moved instead.
  • The animation patterns your agents copy from now ship with the video each one produces, and are grouped by the question you are asking — how one scene becomes the next, how a line of copy arrives, and how a post you already approved turns into motion.
  • A carousel you approved can become a video in five presentations: a flat stack, a 3D coverflow, a rolodex, a triptych, and a receding corridor.
  • A wall of your published posts can scroll as a seamless loop, ending on exactly the frame it started on.
  • Text arrives four new ways: typed at a real rhythm instead of a metronome, one word at a time, line by line from behind a mask, or with a single word snapping into focus.
  • A video can no longer end on seconds of black — its length is checked against the scenes that actually fill it, and a mismatch is refused before it is rendered.
  • A transition written the wrong way used to pass every check and then kill the render; it is now stopped as the code is written, with the line that fixes it.
  • Graphics built from real type are now checked the moment they are made, not only when an agent edits them afterwards — and a first attempt with unreadable or off-canvas text is fixed once automatically before you ever see it.
  • The agent that writes a graphic now works to the same measured rules the rest of the product follows: minimum type size, page margin, line length, colour count, and the brand logo when you have one.
  • A post with no image or video can no longer be approved, scheduled or published — it is stopped with a plain explanation instead of going out empty.
  • Graphics are slower to fail and quicker to recover: a stuck request now gives up in two minutes instead of hanging, and a brief hiccup is retried once instead of losing the whole composition.
  • Every graphic keeps its editable source, so “make the headline shorter” stays an edit instead of a fresh start — and a save that only half works now says so instead of pretending.
  • The 0–10 score on your posts is paused. It was judging still images by a checklist written for video, so almost everything came back as a fail; your past scores are all still there, and you can still ask for a review whenever you want one.
  • Image quality checks now record how many attempts a picture actually took, and say when the limit was reached rather than presenting the last try as the best one.
  • Agents that build motion videos now open the craft guide as a file and are refused until they have read it, instead of carrying it in their head every turn — so a turn that is not about video no longer pays for it, and one that is cannot skip it.
  • The craft your agents follow — the motion recipes, the animation library and the full Remotion documentation — is now a folder you can open and correct, and your edits win over ours without waiting for a release.
  • Keys and passwords are now stripped from an agent’s saved working log before it is written down, and a log is only visible to the person whose conversation it came from.
  • The reference material your agents read is now visible to you as files, and the ones you have edited yourself always win over ours.
  • When an agent hands a job to a helper, it can now open the helper’s full working log instead of taking its word for it — so a job that went wrong can be diagnosed rather than redone.
  • Every new motion video now starts as real scenes instead of one long timeline, so a scene can be rewritten, moved or made longer without breaking the ones around it.
  • A scene whose animation is already over by the time it appears on screen — the reason some videos looked frozen at the cut — is now caught and fixed before the video is finished.
  • Kinetic titles now bring words in one after another at a pace worked out from the words themselves, and the video is lengthened rather than the words rushed.
  • A motion video that is not good enough yet is picked back up on its own until it is — and the agent stops and tells you when two rounds in a row stop improving it, or when it has spent as much on one video as a video is worth.
  • A scene that freezes on screen is now caught before the video is made, on every video — the check used to miss the way most of them are written.
  • A motion video that would come out silent is flagged before it is rendered, instead of after you watch it.
  • The thinking level you pick now really changes how hard the agent thinks — Auto and Pro no longer reason at the same depth.
  • Agents stop re-reading the same brand data twice in one turn, so a turn finishes in about half the time.
  • A graphic can no longer be saved with type too small to read in the feed — the size is checked against the canvas first.
  • When a graphic uses a colour that is not in your brand palette, or type sizes that all look alike, the agent is told before you have to spot it.
  • Before a motion video is rendered, the agent draws a storyboard of it — one picture per scene — looks at each one, and fixes what does not work before spending the render.
  • Motion videos in chat are now checked frame by frame while they are being built, instead of only after the finished file exists.
  • The agents that build motion videos now carry the full library of transition recipes, so the moves the quality check asks for are ones they have actually been shown.
  • Agents that build motion videos now look up the current documentation of the animation library before using something they are unsure of, instead of writing it from memory.
  • Agents get on with the work instead of asking whether they should, or which of two ways you prefer — they only stop for something that only you can answer.
  • A goal now only lists what you actually asked for: an agent no longer turns its own working habits into tasks you have to wait on.
  • Work an agent describes as finished now actually ticks the task off, and a run that ticks nothing gets one informed retry before the goal comes back to you.
  • A goal now says which tasks were ticked off in that very run, not just how many are done in total.
  • A goal in a chat now says how many tasks are left; open it and you see every task, which ones are done, and why a run stopped.
  • On a phone, the menu button inside a panel — or in Settings — now opens the list of sections for that panel, with a way back to the team at the top.
  • The Status button in the top bar is now called Overview, and Overview opens from the Spaces panel too, at the top of the list.
  • Every brand now gets every recurring job on schedule: the ones at the back of the queue go first, so nobody is skipped week after week.
  • Each job on your team page now shows when it last reached your brand, and flags the ones that are running late.
  • The homepage now shows the team at work instead of explaining it in five steps.
  • Blog opens again for brands that already have articles — the page used to sit on a loading skeleton forever.
  • A panel that fails to load now says so and offers a way out, instead of pulsing forever with no message.
  • When an agent needs an app you have not connected, it now asks you to connect it right there instead of saying it cannot help — any app in the catalogue, at any moment, not just during setup.
  • What your agents learn in a chat now becomes part of what your brand knows: a fact you confirm more than once stops living inside that one conversation.
  • Brand knowledge is tidied up every night — what nobody uses fades out, and nothing is deleted before it has been shown to an agent at least once.
  • While an agent is working, its avatar has a life of its own: it changes expression at its own pace and, every so often, pulls a full spin.
  • The avatar no longer freezes into a blank face while it works — expressions morph the way they were meant to.
  • Rows like 3 actions taken stay in the middle of the chat while the answer is still coming in, and the answer itself no longer shifts sideways when it finishes.
  • Each agent now remembers how it works best on your brand, privately — while everything it learns about the brand itself is shared with the whole team.
  • In a group chat more than one agent can answer you, one after the other — each speaks in their own voice, and nobody speaks for a colleague.
  • Agents no longer consult each other behind the scenes: if a teammate is in the chat, they take the floor and answer you themselves.
  • A group chat stops when there is nothing left to add, so you get the answers you need and not a round of agreement.
  • The team on the homepage now follows your cursor — three faces above, two centred below, and the ones on the right turned toward the middle.
  • A link that points at something specific now lands on it: open a knowledge document from a chat source and the document opens, not just the page.
  • Panels no longer offer an Open full page button — a page opens where you are working, and stays there.
  • Spanish and French are fully translated again: the media library, the chat, and every moment an agent asks you to connect an app or open a page no longer switch to English.
  • The page you are on now stands out in the sidebar — a coloured pill instead of a faint purple label — and stays visible when the sidebar is collapsed.
  • Sidebar icons, avatars and counters line up on one column, and the section divider that used to hang over nothing is gone.
  • Clicking a knowledge source in chat opens the document in a panel beside the conversation, already scrolled to the part that was quoted.
  • A broken or mistyped link now lands on a real Anomalia page that takes you straight back — to your workspace if you are signed in, to the homepage if you are not.
  • The homepage demo adds three conversations: three agents taking turns in one room, and an agent asking to connect Instagram or Google Drive before it finishes the job.
  • Onboarding now introduces your team and asks who you want to start with — pick a specialist, and your first chat opens with them.
  • The intro screens show the real thing: your five specialists looking around, the channels your work goes out to, the apps it reads from, and the routines that keep running.
  • The team screen now says what it was missing — when no specialist fits the job, you can hire an agent of your own.
  • A new intro screen shows the work done inside the tools you already use: your product sheets read before the article, so the numbers are the real ones.
  • Setting up without a website now asks you to tell it more, in a box big enough for more than one line.
  • New chats start with a specialist instead of a generalist, so you always know whose craft is answering.
  • The agent picker and the To field now offer the five specialists and your own agents; older chats keep the agent they already have.
  • Agents hand work to each other by craft: a job always goes to the specialist who owns it.
  • In a group chat the agents take turns: whoever the message is actually for answers, one at a time, instead of the same agent every time.
  • Each agent now knows it is in a room and who else is in it — so it stops repeating the others and hands a job to the right specialist instead of doing it badly.
  • While a group chat is working, the loading avatar shows the agent writing right now, and changes when the turn passes to someone else.
  • The header of a group chat shows all its members side by side, not just one of them.
  • The idea bank now rotates: agents see a different set of ideas every time, instead of pitching the same two forever.
  • An idea that becomes a real post is marked as used and stops coming back, so the bank always shows what is still open.
  • The disruptive ideas card is gone from the chat; the idea bank has its own page.
  • A video or photo the AI made now plays right in the conversation instead of arriving as a web address you have to open.
  • Anomalia can now show you photos and videos right in the chat — one large, several as a grid, videos playing with controls — for anything that is not a post.
  • The chat no longer dumps a wall of post cards every time an agent looks something up — posts appear only when the agent means to show them to you.
  • A To field above the message box says who a new chat is for: swap the agent, or add more than one for a group chat.
  • The home screen lost its page title and subtitle — it is the chat, and the To field already says who you are writing to.
  • Chat marks the day between messages, and a New messages line shows exactly where you left off.
  • Weekly post production restarts on its own after an interrupted run, instead of staying quietly stuck.
  • When production keeps failing, you now see it in your notifications instead of finding out from an empty calendar.
  • The agent avatar wears the face of the agent you picked — on the home screen and in chat — and morphs into it when you switch.
  • Agents can now message each other. The exchange happens in its own private thread you can read, and the reply comes back into your chat when it lands.
  • Your AI team and your automations are one roster now: each agent lists the routines it runs, with a switch per routine, and writes every run into its own thread you can read over time.
  • Onboarding no longer ends at a paywall: the agent asks which apps you use and offers to connect them right in the chat.
  • An agent can sign in to GitHub for you: it shows a code in the chat, you approve it in your browser, and it gets on with the work.
  • Captions are written against explicit rules — no AI tells, no repeated call-to-action across a batch — and the reviewer now flags them by name.
  • Anomalia learns from your caption edits: the difference between what it wrote and what you kept shapes the next batch.
  • Text rendered inside images is now read letter by letter before a post ships, and carousel slides keep the same visual thread.
  • Motion videos get real transitions — full-frame scales, match cuts, elements that carry across a cut — checked on the actual composition, not just promised.
  • UGC reels are checked for anatomy: wrong hands or extra limbs now block a take.
  • Music sits under the voice-over, voice clips never overlap, and a video is allowed to run longer so the voice fits.
  • Every text on a motion video stays legible: no overlaps, real contrast over images and footage.
  • Message actions moved into one button that appears when you hover a message, with the model, timing and tokens inside.
  • Tool calls, sources and goal updates are now quiet one-line summaries you can click open, instead of walls of chips.
  • Competitor ads you find in the library now play, and an agent can study one and turn it into a brief — then shoot your version with your products and your people.
  • The sidebar is shorter and says where it takes you: Social media opens the calendar, Web opens the blog, Leads opens the finder — and Ads and Designer moved into the panel.
  • Any page — calendar, content, agents — opens as a panel over what you are doing, so you never lose your place; the page underneath stays exactly where it was.
  • Settings open as a panel over your work; refreshing or opening the link still gives you the full page.
  • Chat text is easier to read: better letter spacing, line height and line length, and the same formatting everywhere.
  • The loading avatar now morphs between expressions and looks around while it works.
  • The brand you are working on lives in one row at the bottom of the sidebar, with brand switching, your account and settings inside it.
  • The message box now sits inline with the controls and only grows when you write more than a line; the context meter is gone.
  • Starting a new chat is instant, the reply no longer flashes twice when it finishes, and the page stops yanking you back to the bottom while you are reading.
  • The chat reads like a real conversation now: replies sit in their own bubble with the agent face beside them, the column runs full width, and long answers stay easy to read.
  • Purple text and badges are darker where they sit on light backgrounds, so they pass contrast at a glance.
  • A motion video can no longer end mid-sentence: the voice-over is checked against the video length before anything renders, and the video is stretched to fit rather than the voice cut.
  • When a motion video studies a reference, it now actually watches the frames instead of reading a summary — and a video with no real transitions, or with a scene that sits still, no longer passes review.
  • Press ⌘K to search everything at once — pages, settings, agents, conversations and the words inside them — and open it without leaving where you are.
  • Keyboard shortcuts throughout: press ? to see them all, and g plus a letter to jump between sections.
  • The bell now counts only what you have not seen yet, and clears when you open it.
  • Conversations now group by day, and a line marks where the new messages start.
  • Creating, editing, pausing or deleting a routine now leaves a line in the chat — click it to see the cadence, the next run and the full brief.
  • Unread conversations now show how many replies are waiting, not just a dot.
  • A live conversation shows a green dot on the agent avatar, the way a chat app does.
August 22, 2026

A chat that never goes quiet

  • Replies generated in the background now show their live progress — the text and the tools as they happen — instead of a silent "Thinking".
  • If a reply dies mid-work, the conversation says so and offers a Retry button. No more threads stuck on "Thinking" forever.
  • You can keep typing while the AI is working: your message reaches it right away, mid-task, instead of waiting for the turn to finish.
  • In Auto, production requests — posts, images, videos, motion, UGC — now run on the most capable model for that turn, and the reply shows which model actually ran.
  • The onboarding conversation now works towards an explicit goal you can see: study your brand, propose your AI team, then ask what you want to do first.
  • Voice-over editing no longer fails on a wrong audio address — it finds your real recording and uses that instead.
August 21, 2026

A clearer chat while agents work

  • Goal updates at the end of a turn are now a proper card — a progress ring, what is done, what is still open, and whether the agent is picking it back up — instead of a wall of raw text. Older conversations get the card too.
  • While a reply streams, the animated avatar now wears the color of the agent actually answering, in every chat surface.
  • Conversations with work in progress now animate their avatar in the sidebar as well, with the same loop as the chat.
  • The elapsed-time label under a streaming reply now lines up with the message text.
  • Chat text is slightly larger on desktop; mobile stays as it was.
August 21, 2026

Your AI team, faster chat, and a one-field start

  • Meet your AI team: every recurring agent now has its own chat and its own on/off switch.
  • Your team starts working the moment you subscribe — each agent introduces itself in chat.
  • The Strategist and the Analyst now work full turns, like the rest of the team.
  • Chat now runs on a much faster model — quicker first reply, same tools.
  • Onboarding is now a single field: your website. The rest is a conversation.
  • The sidebar is now messaging-style, with your agents at the top.
  • Unread dots show which conversations have something new.
  • Agents now send you a message when work is ready — and you can reply to them.
  • You get a clear heads-up before AI credits run out.
  • Motion agents now actually watch reference clips before writing scenes.
  • Deleting a post now also cancels its scheduled publication, from every screen.
  • Voice-over lines can now be trimmed one by one.
  • When an agent asks you something in chat, the choices are now full-size buttons — and you can dismiss the question if it does not interest you.
August 21, 2026

The agent's computer

  • Every agent chat now has a side panel: who the agent is, what it is working on right now, and its routines.
  • The panel shows the agent's latest report and its recent runs — including the reason when a run was skipped.
  • Agents can be paused or resumed straight from their chat, with the same switch as the Agents page.
  • A video render in progress appears in the panel with its elapsed time.
  • "Open where it's working" jumps to what the agent last touched: the post, the plan, or its home section.
August 21, 2026

Agent colors, balanced audio, and real video review

  • Background music in motion videos works again, on a new music engine — beds arrive as 30-second clips that repeat under longer videos.
  • Composer agents each have their own icon and colour again.
  • Background music now automatically ducks under spoken lines.
  • Overlapping voice lines no longer collide on the timeline.
  • Videos now extend a few seconds rather than cut off a line.
  • Video reviews now actually watch the video, not just the code.
  • Video reference research now finds relevant examples instead of repeating old posts.
August 21, 2026

Idea bank, bold ideas, and video quality upgrades

  • UGC videos now keep the same person and product across every shot.
  • Hiring a recurring agent now needs written confirmation, not a casual yes.
  • New free tool at /tools/agent-team proposes an agent team from your site.
  • Bold strategy ideas now appear in chat as a numbered list.
  • New Idea bank (Brand › Ideas) saves, reuses, and adds your own ideas.
  • Motion video animations now ease in and out, not constant speed.
  • Multi-scene videos now render in parallel, finishing faster.
  • Individual scenes can now be redone without touching the others.
  • The assistant now renders and looks at frames before calling a video done.
  • The assistant can now use files, a terminal, and Python in chat.
  • Files the assistant creates stay attached to the chat, viewable from Motion too.
  • Motion videos now include voice-over and music by default.
  • On-screen text and spoken narration are now different and independent.
  • Interface mockups in videos are now interactive: clicks, typing, results.
  • Multi-scene UGC videos now vary camera framing, not just action.
  • The UGC status panel is now scrollable, like Motion's.
  • Finished motion videos can now be requested straight from chat.
  • Motion videos now include ready-made transitions, shapes, and animated lines.
August 20, 2026

Goals: work continues until it is done

  • Long tasks now start with a visible checklist that updates live.
  • Work continues automatically in the background until every item is done.
  • If it gets stuck, it reports exactly what's missing instead of pretending to finish.
  • New `/goal` command sets, checks, or stops a conversation's goal.
  • Chat commands now open by typing `/`, with filtering and keyboard selection.
  • Composer menus now close by tapping outside or pressing Esc.
August 19, 2026

More honest content, more hook variety

  • Posts are now checked for AI-sounding phrasing, not just accuracy.
  • The AI no longer invents statistics, testimonials, or customer names.
  • New pre-publish caption review checks captions from five angles.
  • Eighteen recognized post-opening styles, up from seven, cut repetition.
  • Brand brief now includes what your product does NOT do and banned claims.
August 19, 2026

Clearer limits on what analytics can prove

  • Weekly recaps no longer declare winners when the sample size is too small.
  • Every analysis now states what it could not determine, and why.
  • Paid campaign issues are now diagnosed by cause, not just flagged.
  • Quality scores now separate real issues from checks that could not run.
  • GEO scoring now covers five factors in how AI models cite your brand.
July 21, 2026

Carousels and in-chat previews

  • The AI assistant can now create multi-slide carousels, not just single images.
  • New posts now preview directly in the chat conversation.
  • Content grid now labels carousels and videos at a glance.
  • Carousel posts can now be browsed slide by slide.
  • The post edit window now always opens centered on screen.
July 21, 2026

Radar: new lead sources

  • Radar now also searches Threads and LinkedIn, not just Reddit and news.
  • Radar finds opportunities even without a connected profile on that platform.
  • LinkedIn radar surfaces posts where your expertise can win leads.
July 12, 2026

Blog: categories, authors, search

  • Blog articles can now be organized into categories and tags.
  • Articles can be assigned an author with photo, bio, and profile page.
  • New full-text search added to the public blog.
  • Blog navigation is now customizable, including top bar or sidebar layout.
  • New AI Humanizer rewrites generated articles to sound more natural.
  • Blog homepage can now use a fixed left sidebar layout.
July 12, 2026

Radar and notifications: multilingual

  • Radar sources now support over 30 languages with automatic detection.
  • Notification emails now go to all brand collaborators, not just the owner.
  • Each collaborator receives notification emails in their own language.
July 12, 2026

Social image format fix

  • Instagram and Facebook post images now use the correct format automatically.
  • The story format was removed since it can't auto-publish to the feed.
July 10, 2026

Blog hosting and custom domain

  • Every brand now gets a dedicated public blog with AI-written, SEO-optimized articles.
  • Blogs now support a custom domain.
  • Articles can now publish to Shopify, Webflow, and Wix too.
  • Articles can now be scheduled for automatic publishing.
  • AI can now plan a full month of article topics and dates.
July 8, 2026

Radar: news and trend monitoring

  • Radar now automatically monitors news, subreddits, and RSS feeds.
  • New daily email digest with the most relevant news and suggested actions.
  • Suggested sources can now be approved or rejected directly from the email.
July 5, 2026

SEO and GEO optimization

  • New SEO analysis with a strategic keyword plan.
  • New GEO audit checks how AI models cite your brand.
  • Articles can now be auto-generated for target keywords with a quality score.
  • Articles can now be auto-optimized to reach a 90+ quality score.
July 1, 2026

Autopilot: automatic content generation

  • Autopilot now generates social posts and blog articles from your weekly plan.
  • Content is generated in your brand voice, tailored to each platform.
  • New weekly email recap of activity and performance.
  • The publishing queue now manages itself, with automatic rollback on errors.