Copera Office Arrives, Agents That Work on Their Own & a Transcriptions Hub

Copera Office lands: Word, Excel, PowerPoint and Markdown files now open and edit inside Copera, with an AI panel that writes straight into the document, an Ask AI button on any selection, every AI edit listed and revertable, and 30 days of version history. Alongside it, hired agents can finally do a job end to end — they have a page listing everything they can reach, can work across several boards, and can be given narrow, revocable authority to email a lead, answer a customer or book a meeting without asking each time.

New Features

Copera Office — Word, Excel, PowerPoint and Markdown, Inside Copera

Office files are no longer something you download, edit elsewhere and upload again. They open in Copera, in a real editor that looks like the rest of the app (#3690).

• Four formats, one editor — Word documents, Excel spreadsheets, PowerPoint presentations and Markdown files open natively, with co-editing, presence avatars, read-only mode when you only have view access, and a save status you can trust. The editor chrome follows your theme and the area accent, so it stops looking like a guest in the window.
• An AI panel inside the document — ask it to write, rewrite, translate or insert something and it puts the result into the file, placed by your selection: replace to transform, insert after to add, or at the end. It reads the open document before answering, confirms in one short sentence instead of pasting the content back at you, and follows Word's Track Changes switch so its work arrives as a suggestion when that is on.
• Long writing streams into the page — ask for a long piece of text in a Word file and you watch it being written as a preview outside the document. If it stops early you get Keep or Discard, and Keep never duplicates text that already landed.
• Ask AI on a selection — select text and a small mark appears at the end of it. Click it (or press Ctrl/Cmd+J) for Rewrite, Make shorter, Make longer, Fix spelling & grammar, Change tone, Translate to any of the six app languages, Summarize, or a free field. Readers are offered Summarize only. It works on Excel ranges and PowerPoint shapes too.
• AI changes, listed and revertable — a panel shows every edit AI Chat or an agent made to the file, with what it touched (paragraphs, cell ranges, slides, sheets). Revert one, revert everything, or revert the rest and keep what people changed since. Older changes move into version history.
• Version history — every save adds a version, kept for 30 days and never fewer than the last 20. Restoring one saves it as a new version, so nothing is lost, and versions say who saved them and which were AI-authored.
• CSV and TSV files open as spreadsheets — with a clear badge saying that saving keeps values only, and exactly what is not saved (sheets after the first, formulas, formatting). Save a copy as Excel workbook gets you a real workbook next to it when you need the rest.
• Turn a PDF into a document, spreadsheet or presentation — convert a PDF from Drive or ask AI Chat to do it. Pages with no text are kept as images, and you are told plainly when a PDF is password-protected, damaged or too complex to convert.
• The editors speak six languages — French, German and Italian join English, Spanish and Portuguese across the Word, Excel and PowerPoint interfaces, including the Word table controls.
• Fonts and numbers that match Excel — a cell with no font of its own is now drawn and measured in the workbook's normal font, so columns no longer show ##### or render in the wrong face on machines without Microsoft's fonts.

AI Agents — They Can Now Work on Their Own, Within Limits You Set

A week of making hired agents actually finish the work: support, sales and marketing jobs that run to completion, and the controls to let them (#3710, #3713, #3715, #3725).

• An Access page per agent — everything the agent can open, by name, grouped by kind: boards and tables, pages, Drive folders, channels, inboxes, Omni channels, calendars, contact lists. Add access lets you pick several items at once, each at its own level, and shows you who gains what before anything is written. Members see it read-only; admins can remove.
• One agent, several boards — an agent can ask for more than one board in a single card, and you can grant all of them in one answer. Giving it a second board no longer quietly breaks the first.
• Standing authority — grant an agent narrow, revocable permission to act unattended: send one first-touch email per lead in a board table, reply in-thread to people who wrote to an inbox, or book meetings on your own calendar with people from that inbox. Each grant has a daily cap, a per-thread cap, an end date, and an owner who can revoke it at any moment. Only the inbox's owner or an administrator can grant the inbox ones, and only you can grant access to your own calendar. Everything outside the grant still asks first.
• Support that behaves like support — answers are written in the agent's voice and the customer's language, a message with two questions gets both answered, images and voice notes without a caption are understood instead of escalated, bug reports are written once into the table bound at hire with a customer-facing reference number, and the customer is told when a person takes over. An agent stands down on a thread a person is handling, and the thread can be handed back to it later.
• Sales that keeps its promises — a reply claiming real work ("I emailed Ada", "booked Tuesday 3pm") is only kept when there is a receipt behind it. Meetings with an outside attendee email the invitation and any reschedule notice by default, and a lead is marked Meeting scheduled only after the calendar entry actually exists.
• Marketing that publishes — the drafting step can search the web and read pages, and must cite what it read. Publishing writes the live URL and marks the row Published, replays safely after a crash, and never publishes the same row twice. Optional auto-publish is off by default and capped at three posts a day.
• Reports stay private — a routine's report to its owner is delivered privately, with only a neutral line in the shared channel, and a replay never drops a private report into a channel someone has joined since. The same rule covers cards, step titles and anything the agent might repeat afterwards.
• Routines tell the truth — a run that did not finish its goal ends incomplete and says what is missing, a routine told to research must actually research, a stated row count it had no tool for is reported as a shortfall, and the owner's report opens with what the run really did (rows, statuses, files, failures).
• Ask Copera finishes the setup — ask for a whole setup and it builds the board, the agent, the automations and the access in one reviewable plan. A plan that fails part-way keeps what it made, so the next attempt completes the setup instead of rebuilding it. A name already taken asks you for a different one instead of taking over the existing agent, approval cards name the board, table, column and status they act on, and Run now answers immediately and follows the run to its outcome.
• Routines and agents can be set up from an inbox — Ask can create an automation scoped to a single inbox, so an arriving email wakes the agent you chose.

Transcriptions — A Hub, Automatic Summaries and Email

Every recording Copera transcribes now lives in one place, with a summary waiting for you (#3720, #3727).

• A Transcriptions hub — a drawer with a list and a preview, All / Mine / Shared, a filter by source and a search. Phone recordings are their own source.
• A summary for every transcription — summaries are now produced by default for all transcriptions, not only meetings, and the hub and preview use one rule for whether a summary exists yet.
• Email the summary and the transcript — pick who gets it, add a note, and send the summary in the email body with the full transcript attached as a PDF, each recipient in their own language. There are clear limits and clear refusals: who is outside your workspace, which address is invalid, what is not ready yet.
• A page for phone recordings — recordings made from the phone app open on their own page with the viewer, AI actions, rename, share, delete and email. New phone recordings trigger the same summary and automations as everything else, a silent recording reads as "No speech detected" rather than failing mysteriously, and phone memos stay private unless you deliberately share them.
• Create a page from a transcription — one action turns a transcription into a page, and you can ask AI Chat about any transcription from the viewer.

Shared Inbox Invitations on Everyone's Calendar

An invitation emailed to a shared inbox used to be one person's problem. Now it is the team's (#3723, #3724).

• One event, all the members — an invitation to a shared inbox appears on the calendar of every current member, with its own state: watching, attending, handled, declined or hidden.
• One team answer — the reply is sent once, from the invited address, and I'll attend lets a single member take it on. Attendee-only reminders and calendar copies mean nobody is pinged for a meeting they are not going to.
• Per-member and per-inbox preferences — plus a transferable inbox owner, so the person who holds an inbox can hand it over. The inbox owner can't simply leave; ownership moves first.
• Remove it from your calendar alone — "Remove from my calendar" changes only your view; the other members keep the invitation.

Boards — Automations That React to Lookup Columns

Automations could not see a lookup column at all: a change in the mirrored values produced no event, and row conditions compared the wrong value (#3711).

• Lookup changes now wake automations — a trigger or event routine that names that exact lookup column runs when its values change.
• Conditions and filters read the mirrored values, so "when the client's plan becomes Enterprise" finally works from a lookup.
• Values stay hidden from people who shouldn't see them — run history, approvals, notifications and every reading surface go through the same check against who is actually reading, and fail closed if the lookup was deleted or re-pointed.

Desktop — Smarter Tabs, Real Pickers and a Launch Screen

• The + menu lists what you actually use — Channel, Direct message, AI Chat, Board, Dashboard, Canvas, Page and File from Drive each open with your most recent items already listed, 20 at a time, paging as you scroll or with Arrow Down. Typing filters everything; pages, files and the search across everything query the server. Boards, dashboards and canvases are ordered by when you last opened them. The create column gained PDF, AI chats, dashboards and canvases (#3690).
• A launch screen instead of a black window — a Copera rocket idles on its pad under a status line that rotates through the six app languages, and lifts off when the app takes over. It respects reduced-motion settings, remembers your theme and language for the next launch, and the web app shows the same screen (#3690).
• Rename a file from its tab — double-click a Drive file tab (or press F2) to rename it inline. The tab, window and editor titles all update at once (#3690).
• Another workspace opens in its own window — with desktop tabs on, opening a second workspace used to replace the one you were in. It now asks the desktop app for a real window (#3722).
• Dialogs and toasts keep clear of the window controls — Tempo, board permissions and full-screen surfaces no longer sit under the minimize and close buttons on Windows and Linux, and the top bar stays flush with the top of the window (#3717, #3690).
• The top bar is in a sensible order — account, bell, mentions, threads, calendar, Spark, meeting bot, AI Notes, then the Bankai star, which is now present at every window size, has a tooltip and toggles the assistant open and closed. When icons fold into the overflow button, that menu is a proper list with each item's name beside it (#3690, #3708).

Voice Messages — A Real Waveform, Pause and Resume

• A proper player — voice messages get a drawn waveform, reliable clicks and keyboard control, and tighter spacing in the message row (#3716, #3719).
• Pause and resume while recording — with clear states for recording, paused and resumed, and a sound when a recording completes (#3716).
• Your recording survives a routine sign-in refresh — a queued upload used to be tied to the exact session token, so a background token refresh made your own upload look like someone else's and it was silently deleted. Uploads are now tied to you, retry once by themselves, and a final failure keeps the recording with a Retry button instead of dropping it (#3730).

Copera in Your Browser — Runs That Actually Work

Almost every browser step the agent took was failing in production. That is fixed, and the setup is clearer (#3699, #3700, #3703).

• No more hangs — a screenshot that Chrome never answers gives up after five seconds and retries once instead of freezing the rest of the run with "Previous action is still unwinding."
• Screenshots you can actually read — the black box that covered the top-left quarter (or the whole image on a standard screen) is gone, and images are now the right size on high-resolution displays, so clicks land where they were aimed.
• It works while the window sits behind other apps — clicks, typing and screenshots no longer need the browser window in front.
• Stop means stop — stopping from the run viewer, from the Copera pill or from the chat all end the run, and the agent is told the stop is final for that turn instead of starting the same task over.
• Connect steps are real steps — numbered steps with a title and a hint each, the Chrome Web Store as an actual button, and an outdated extension gets update steps instead of install steps that contradicted the warning (#3699).

Bug Fixes

• A status change could skip its required fields — a transition's required fields were only checked by the dialog, so a row whose fields an automation had cleared could move forward again from a stale tab, and anything writing outside the dialog skipped the check entirely. Required fields are now enforced wherever a person writes, and a required field you cannot see says so instead of naming it (#3721).
• Typed names rewritten in the sidebar — every sidebar item was capitalized, and because a dot counts as a word break, Recents showed "Budget.Xlsx" and board, channel and doc names were rewritten. Names now appear exactly as you typed them, in the sidebar, the voice chat header and the Drive details panel (#3721).
• The Office editor loading in the left 40% of the window — the editor now takes the full width from the first paint and shows a per-format placeholder while the file loads, and the save status no longer reads the opening handshake as an unsaved change (#3721).
• People unable to send in direct messages they did not start — a permission check added last week required a role flag that only a channel's creator ever had, so every DM recipient, group-DM member, voice, meeting and classroom participant got "Message destination is unavailable". Membership is now decided by who actually participates, including inside threads, while viewers stay read-only (#3701).
• Shared team row assignments disappearing — rows assigned through a team, their activity, workload labels and My Work entries are all restored, and removed assignees are refreshed properly (#3712).
• Guests blocked from boards they were invited to — guests can read their own board permissions and the row templates they have access to, and row visibility is scoped to the table being asked about.
• Shared inbox invitations leaking into members' personal calendars — the leaked copies are cleaned up, a daily sweep catches any that were missed, and replying in a thread reads the attendees from the email that was actually answered (#3714).
• Duplicate events left behind in Google Calendar — a crash during a sync could leave an untracked copy nobody could clean up, a member leaving a shared inbox could leave their copy behind, and a cancelled occurrence could still be pushed. Copies are now recorded before they are created, reconciled by a regular sweep, and removed when the person is no longer entitled to them (#3726, #3728).
• A declined invitation coming back as accepted — the calendar catch-up work could re-attend an invitation somebody had already declined in the Inbox. A member's own choice now always wins, and anything still undecided is left to settle on its own rather than being guessed (#3731).
• Inbox labels applied to the wrong conversations — contextual label classification is more accurate and keeps its decisions when a retry happens (#3706).
• The unassigned-conversation alert telling human teams an AI was unavailable — the SLA alert reused the AI handoff headline, so a human-only support queue was told a conversation needed a human "because no agent was available". It now simply says the conversation is waiting for the team, in all six languages (#3711).
• Reactions on row activity not appearing until a reload — a reaction now shows the moment it is confirmed (#3711).
• A new direct message's duplicate-conversation note painted over by history — the header no longer shrinks below its content, so the note and its action wrap inside it (#3711).
• Drive and channel polish — share avatars no longer clip, Drive columns resize, Workspace Disk sharing matches the disk's own rule, avatar stacks have hover cards everywhere, the grid view scrolls sideways over blank area, and Drive files open as tabs on desktop (#3714).
• Office tabs cut off at the bottom — inside a desktop tab the editor ran a title bar past the window, so the AI panel's composer and a scrollbar were clipped. The editor now fits its host box. Word, Excel and PowerPoint also no longer borrow each other's stylesheets, which had been laying the Word toolbar out as a spreadsheet ribbon (#3690).
• The first AI edit of an Office file always refused as out of date — a document nobody had open was read with one set of internal ids and edited against another, so every first edit missed. Both now agree (#3690).
• AI panel edits claiming work they had not done — "Rewrote the selected paragraph" with no edit behind it is now caught, and an AI answer no longer treats the record of earlier tool work as proof that this request is already finished (#3690).
• The AI answering a translation request in the wrong language — asking for a translation now changes the file's language and never the reply's, and chat titles are written in the language the message was written in rather than one it merely names (#3690).
• Ask Copera publishing a to-do plan in the wrong language — a continuation could publish its plan in Spanish inside an English conversation. The plan is now checked against your own words, and a language code or name you asked for is respected — including a board named something like "PT-BR Launch", which is a name, not a request (#3729).
• AI answers abandoned when a provider stalled — a hung provider call now times out and moves to another candidate instead of spinning until the run's deadline, and a step that hits a capacity limit after work has already been committed carries on elsewhere rather than failing the whole turn (#3690, #3709).
• Moving a row between tables losing every cell — a data-loss bug found by driving a real request through Ask Copera: the row survived the move but arrived empty. Fixed, and retained-image access follows the row to its destination (#3713).
• A non-admin unable to duplicate a row on a board with a restricted column — duplication was filtering restricted cells against the wrong thing, so it matched nothing and failed validation (#3690).
• Automation reliability across AI and Spark — a batch of fixes to automation runs, AI handling and Spark transcoding, including cleanup of abandoned work and search indexing of contacts (#3705).
• Linked-row automations, guest roles and row access — linked-row automations are more reliable, assignee eligibility is checked properly, and invite resends respect board authority (#3704).
• Mobile notifications going missing — pushes now retry properly when delivery fails, your notification tone preference is honoured, and rich notification content is rebuilt correctly for older app versions (#3718, #3716).
• Voice calls on older app versions — legacy call clients are protected again, and call dismissal timing is fixed (#3716).
• Board activity, grid and permission edge cases — board workflow policies refresh when a column changes, comment notifications resolve team assignees correctly, and row-security rules with a "not equal" condition now hide and show the right rows (#3712, #3690).

Improvements

• Your AI allowance goes all the way to 100% — hidden ceilings below the limit you can see are gone. A request is refused only when nothing is left, and near the end the answer is simply sized to what the remainder can pay for rather than being rejected. A finished answer is always delivered, even if its receipt is lost, and the agent header shows "under 1%" instead of rounding real spend down to zero (#3697).
• New agents start with every power and narrow themselves — a new agent begins with full access and is told to give up what its job does not need, with the safety floor intact: it still asks before sending, posting, buying, deleting or signing in anywhere. Widening a power still takes an admin (#3697).
• Long AI jobs get the effort they need — the assistant now picks how hard to think based on the work in front of it and widens its time budget when a job turns into a long run, so a big piece of work finishes instead of stopping half-way (#3702).
• Automations can trigger every meeting AI action once the transcription is ready — with a note in the builder spelling out what downstream blocks can do with the finished meeting (#3727).
• Rich text written through the Public API and AI goes through the same path as every other column — so a workflow's required fields see it in the same call, writes are not applied unless the row write actually committed, and clearing a cell with an empty value works as documented (#3732).
• Row numbers are reserved properly — duplicating rows, recurring to-dos, imports and board setup all reserve their numbers atomically, so two rows can never claim the same one (#3715).
• Drive's new menu is clearer — Document, Spreadsheet, Presentation, Markdown file, PDF, folder, upload, Record a Spark and Send for signature, with a line telling you which folder new files are saved in (#3690).
• Drive Home suggests what to create — per-format prompts ("a Q1 budget based on Q4 with 5% growth", "10 slides from the Investor Day notes") and a label on files that were drafted for you (#3690).
• DocSign takes any Office file — choose a Word, Excel or PowerPoint file and Copera prepares the PDF for signing, saving it next to the original in Drive. Drafts now have their own box in the DocSign inbox (#3690).
• Agent run viewers show more — which browsing model ran, whether it could see the page, the extension version, and "Stopped early" with the reason instead of a misleading "done" (#3707).
• A browser run that ends early keeps what it found — a run cut short by a step limit or an unclear provider reason is marked incomplete and keeps its findings; only a run with genuinely nothing to report fails (#3707).
• Agent browsing waits for the page — the extension settles after each action that changes the page, so an asynchronous click or a lazy scroll is no longer read as "nothing happened", and a vision-capable model gets a screenshot after each action so it can see what it did (#3710, #3707).
• The command palette is hidden from guests, and its requests wait until it is actually needed.
• Office files never commit a version identical to the last one, and version history for CSV and TSV files appears only where the spreadsheet editor is available (#3690).