Agents Get Their Own Channels, a Vault for Secrets & Copera in Your Browser

The biggest change to Copera's AI yet: an agent is now a channel. You talk to it where you talk to your team, it asks for what it needs right there, and it answers to its own admins with its own powers, budget and memory. Next to it, the Vault means a password never has to sit in a message again, Ask Copera can now do almost anything you can do in the app — and asks first when it matters — and a new Chrome extension lets Copera work in your browser alongside you.

New Features

AI Agents — An Agent Is a Channel You Talk To

Hiring an agent used to mean a profile page and a direct message. Now every agent is a channel in your sidebar, with its own face, its own members and its own settings (#3650, #3675).

• Create one from the create-channel dialog — pick a bot (four characters in twelve colours), name it, choose who has access and set how much of the week it may spend. The eyes are its status display: thinking, working, waiting on someone, paused, failed, or a brief tick when it finishes. There is also a New Agent card on the Communicate landing page.
• Every message in that channel is for the agent — no more wondering whether to mention it. Threads under the channel work the same way, and turns run one at a time so two questions never collide.
• It asks in the channel and continues from your answer — a permission it needs, a resource to pick, a secret to enter or a plan to approve all arrive as a card in the conversation. Allow once, always allow (bound to that agent, tool and resource), or deny. After you decide, the agent picks up exactly where it stopped.
• An amber badge when it needs you — the agent's row in the sidebar counts the requests you can actually answer, and a slim notice above the composer says when it is paused, out of budget or stuck, with a way to resume it.
• Powers instead of a capability list — 26 readable powers with levels (off, view, edit, full) and presets, so you grant "Boards at edit" rather than picking through individual abilities. Widening powers needs an agent admin; narrowing them does not.
• A budget as a share of the week — set an agent's weekly limit as a percentage of your workspace's AI allowance, never as a money figure, and see where the week went, split across the people who asked. Routines count as their own asker.
• A brain page — its instructions with a version history you can restore, and its memory by topic, searchable, with forget-one and forget-everything.
• Settings that look like channel settings — Overview, Powers, Members, Budget, Brain, Routines, Notifications and a Danger Zone. Members shows what each person gains through the agent before you apply it, and the last admin cannot leave.
• Agents answer to their own admins — administering the workspace does not make you an admin of every agent. Creating agents, deleting any agent and administering all agents are three new permissions in the role editor, the last of which only the workspace owner can grant.
• Workforce becomes the roster — one page listing the agents you can see, with New agent right there.
• It sets itself up — an agent can update its own profile, write its own instructions, ask for the powers it needs and narrow them again, all inside the rules you set.

The Vault — A Password Never Has to Sit in a Message

A new vault keeps secrets out of your conversations, out of Copera's search index, and out of any AI model's context (#3650).

• Seal from the composer — a lock button next to send. Name the secret, type it in the popover (never in the editor, so it never reaches a saved draft), and the message carries a chip with the name instead of the value.
• Opening a chip asks you to prove it is you — being signed in is not enough. Unlock with a passkey, a PIN, an authenticator code, your phone's biometrics, or by approving on your phone with number matching. The value is shown masked from session recording and disappears on the server's timer.
• A vault per person and a vault per agent — My vault lives in Settings → Security. An agent's vault is opened by its members after they prove who they are, admins-only items stay closed, and deleting an agent can move its secrets somewhere else first.
• "This was a secret" — already sent a password in the clear? One action rewrites every copy Copera holds: the message, its translations and transcription, scheduled copies, the search index and anything the AI distilled from it. It can turn the text into a vault chip, and it always tells you plainly what Copera cannot reach, so you know what still needs changing.
• A nudge, not a wall — if a message looks like it carries a password, a slim bar offers Seal it or Send anyway. It asks once per message, so nothing gets trapped behind it, and it only appears where sealing is actually offered (#3684, #3682).
• Agents can use a secret without ever seeing it — an agent passes a reference, and the value is decrypted at the moment of the request, only for a site the secret was sealed for, and scrubbed out of the response before the agent reads it.
• Forgot your PIN? — the locked card offers a reset by email. The link does nothing on arrival, so a mail scanner cannot wipe anyone's PIN; you confirm with a click, and the vault reopens after a cool-off.
• Rename an item in place — only the name. A value is never edited; you seal a new secret instead.

Ask Copera — It Can Now Do Almost Anything You Can Do

Ask Copera's abilities are loaded in packs that stay out of the way until they are needed, so it can reach far more of the app without slowing every question down (#3650).

• Eleven new areas — channels and categories, docs, members and teams, calendar and bookings, email, Omni conversations, DocSign, board administration, Drive and templates, connections, and budgets. Every one of them acts as you, through the same services the app uses, so nothing an agent does bypasses a permission you do not have.
• It asks first where it matters — sending anything outside your workspace, deleting, sharing with a guest, publishing a form, or anything that would cost money always shows you a card, in every permission mode.
• Build plans — ask for something bigger ("set up a board, a channel and an agent for onboarding") and you get one card describing every step in order, with the strictest rule of any step applied to the whole thing. A plan that fails halfway stops there, and retrying does not repeat what already landed.
• Always allow, for real — a standing grant can now be tied to the resource it was given for, so allowing rows in one board says nothing about another. Settings gains an Allowed actions page where you revoke your own grants, and administrators the workspace-wide ones. Delete-class actions and publishing an automation can still never be always allowed.
• Paste text, HTML or JSON as a source — a long paste or an HTML, JSON or text file now becomes a proper source attached to the conversation rather than a wall of text in your message, and it stays a source through edits, forks and history (#3657).

Copera in Your Browser — A Chrome Extension

A new Copera extension for Chrome lets the assistant work in a browser tab with you (#3648, #3662, #3664, #3666).

• It sees the page and acts on it — open tabs, navigate, read the page, click, type and take screenshots, with a cursor drawn so you can follow along.
• It clicks what you see — clicks can be aimed at a point on a screenshot, and a screenshot is thrown away after anything that changes the page, so it can never act on a stale picture.
• It knows when it is getting nowhere — the same no-effect action is refused the third time instead of looping, and the assistant is told to stop and report instead.
• Sensitive pages are fenced off — private addresses are blocked, password and card fields are treated as protected, login submissions are held back, and field values are masked out of everything the assistant reads.
• Published on the Chrome Web Store — version 1.0.0 with proper toolbar icons and store listings in English, Spanish and Brazilian Portuguese (#3666).

Inbox — Label Pills, Automatic Labels and an Inbox AI You Can Set Up

The Inbox gets its folder bar, its labels and its AI settings redesigned (#3681, #3674, #3670, #3685).

• Folder pills — system folders first, then your labels, then Create label, each showing its name on hover and folding into a More menu when space runs out. A carved divider separates system folders from labels.
• Labels where you expect them — label chips sit next to the subject in an opened email and the Labels control moved into the top toolbar, so the empty row under the subject is gone.
• Automatic labels with clearer rules — "How do emails get this label?" now spells out the choices and suggests rules, and an opt-in box labels the 500 most recent emails you already have. If that pass fails, your label is kept and Retry repeats only the labelling.
• Turning automation off stops it — the backfill re-reads your settings between chunks, and a label you change by hand stays changed. Removing an automatic label shows a short Undo message and tells you Copera will not add it back to that conversation; Use automation again in the Labels menu hands it back to the rule.
• Summaries — a single email gets a "Summarize this email" button; a conversation with two or more messages summarizes when you open it.
• Inbox AI, not "Powers" — a first-time setup checklist and a grouped settings panel that spells out which features depend on which, and saves the features you picked in the same step that turns it on.
• Row actions that sit in the row — hover actions lost their box, each icon gets its own round hover, and a long sender name shortens with an ellipsis instead of pushing them off.
• Desktop motion matches mobile — row animations, folder tabs and undo behave the same on both, and trash undo on desktop is reliable (#3673).

Boards — Pick the Header Row When You Import a CSV

An export that opens with a merged title banner used to make the real headers a data row, rename your blank cells and drop a column (#3669).

• Auto-detected, and overridable — the wizard finds the header row for you and offers a Header row selector over the first 100 rows to override it.
• Blank and duplicate headers are counted on the real header row, so the warnings finally match what you are importing.
• Importing into an existing table no longer dead-ends — an import whose headers match nothing can continue as long as you set a type on the columns to create (#3668).

Meetings — Stop the Music

The shared music player in a voice channel gains a proper stop that ends playback for everyone and clears the queue, with a confirmation when there are still tracks waiting. Skip is disabled when the queue is empty, the transport controls stay visible on the music tile instead of hiding when your pointer leaves it, and the button reads Play with a play icon while nothing is playing.

Bug Fixes

• The AI answering in the wrong language — an English request from someone with a Brazilian name, location and time zone produced a Portuguese board, Portuguese rows and a Portuguese reply. The language rule now comes last in every prompt and covers everything the assistant creates, not just its reply: an explicitly requested language wins, otherwise the language you wrote in, and never your location, name, time zone or stored data. Summaries, extraction, deep research, emails, voice modes, automations and dashboards all follow it (#3672).
• AI turns failing outright whenever email was involved — any turn whose abilities included sending an email, inviting someone, booking a meeting or requesting a signature was rejected before it started on several models, because of how the email check was described to them. Fixed for every model (#3691).
• Ask Copera showing a dead end instead of a way forward — a permission problem inside a turn used to kill it with a bare card. The assistant now receives the refusal and can retarget or pick an allowed tool, and a turn that legitimately ran hundreds of steps no longer fails at the finish line while its work sits completed. A failed turn now offers Continue next to Retry, says how many steps finished before the stop, and Report problem actually sends a report instead of copying an internal reference to your clipboard (#3654).
• Plan-mode conversations failing to load — every Ask Copera conversation in plan mode came back malformed, and receipts carrying context information were rejected whole, which lost the cost row. Plan steps that failed, were cancelled or are waiting for review are now shown instead of silently dropped (#3693).
• Formulas not following standard math order — unary minus and NOT now bind tightest and AND binds tighter than OR, as in every spreadsheet. Existing formulas were rewritten with the minimum parentheses needed to keep producing exactly the result they produce today, so nothing you have built changes value (#3692).
• Row writes failing on boards with a filtered link, lookup or rollup column — a filter set up before those columns supported filtering made every row create and update on the source table fail. Old filters are read correctly now and stored ones were repaired (#3669).
• Lookup, rollup and formula cells only updating after a reload — recomputed cells on a board with per-viewer row visibility or a restricted table were broadcast blank or not at all. Every place that recomputes now tells the board to refresh, and dashboard widgets pick up a cell edit instead of re-serving an aggregate up to 35 seconds old (#3669, #3656).
• Crashes and stalls in the link and lookup engine — rows saved in an older shape crashed lookups and automation triggers, deleting a row could fail an entire batch, and the nightly job that repairs link and lookup drift had been dying on its first step since 31 August. All three are fixed and the repair job runs again (#3668).
• "Add a new task" vanishing inside an (Empty) group — adding a row into the (Empty) bucket of any grouped grid flashed the row in and lost it. The row now lands in that group and an active filter cannot move it out (#3668).
• Group headers showing a raw identifier — a Status, Dropdown or Labels cell still holding an option you deleted painted its internal id into the group pill, into cell text, into copied data and into CSV exports. A value with no live option now reads as (Empty), and a Labels cell keeps only its live options, so two groups no longer both read "Done" (#3665).
• Board activity history and grouping controls disappearing — activity pages filtered by permission are traversed properly, so history is preserved and the grouping controls stay put (#3689).
• Board templates firing automations on every change — saving a board as a template blanked the target status of a "changed from / to" condition, and since blank now means "any value", every copy of that automation fired on every change of the watched column instead of only on the configured transition (#3659).
• Automation canvas showing a record's id instead of its name — add and remove field-value blocks printed the raw identifier of a linked record on the node; it now shows the record's name (#3669).
• A dialog opened from a menu closing the moment you clicked in it — opening "Manage" from a view's options menu left the menu on top, and clicking the name field dismissed both, so the view could never be edited. Dialogs no longer live or die with the control that opened them, and every menu item that opens one now closes the menu first: manage view, add view, manage table, import, delete column, copy option ids, edit chat category, disable recents, edit signature and new Drive content (#3659).
• The meeting music panel opening behind the menu you opened it from — the same shape affected the whiteboard and document session lists. All three now sit above the menu, which closes as you would expect (#3659).
• The row dialog filling the window and leaving the board fully lit behind it — it gets its top and bottom margin back, the background dims like every other dialog, and clicking outside closes it (#3659).
• Surfaces stacking by accident — the app now has one named layering order instead of magic numbers, so the editor link popover, grid cell editors, bulk-action bars, the route progress bar and Mission Control's overlay all land where they belong, at every window width (#3659).
• The automation run history drawer dragging the whole screen sideways — opening it scrolled the work area, pushed the fullscreen builder left and showed the board through the gap until the animation landed. The same defect in Tempo's overlay drawer is fixed too (#3665).
• Tempo dialogs painted in the wrong palette — booking, time-off and confirmation dialogs opened from the calendar now keep Tempo's own colours (#3659).
• Collaborators missing new rows on boards with restricted row visibility — a board narrowed only by role assignments (a client or contributor preset, or an own-scoped custom role) sent row values it should have held back, and blanked grid updates left collaborators without the new rows and with stale cells. Both are fixed (#3660).
• Long toasts filling the screen and covering a dialog's close button — toast titles and bodies are now clamped to two and four lines with a Show more toggle when text is actually cut off, top toasts drop below the header while a fullscreen dialog is open, and every toast has a close button (#3669).
• Opening an email then coming back to Communicate landing you in the last chat channel — the Inbox was never recorded as somewhere to return to, which also sent post-login restore and the Settings close button to chat. Drive had the same problem: Sparks, Trash, DocSign and office files were forgotten by Collaborate (#3655).
• Standalone row links closing onto nothing — opening a row by its own URL now resolves the board view behind it and closes onto that view.
• Google sign-in not completing on the desktop app — it now opens in your system browser and hands the session back to the waiting app, keeping the page you were heading to.
• Office files too large to save opening in edit mode — a file above the save limit opened as editable and then refused every save. Such files now open view-only for everyone (#3685).
• Two "Sales" inboxes — inbox addresses are saved in lowercase and checked case-insensitively, so Sales and sales can no longer both exist. The address fields show lowercase as you type (#3681, #3685).
• The Settings dialog painting behind other panels — it is placed into the work area properly, and its two-pane layout ignores the default gutter.
• The sign-in page for the command line clipping its content — a tall page scrolls instead of being cut off at the top.
• Dark theme readability in the Inbox — text you select in an opened email is readable, and the "Use automation again" control on a card uses the card's own colours instead of disappearing into it (#3681).
• Agent runs reporting the wrong status, and finished steps losing their outcome — a run's terminal state is settled per run, and a finished step keeps its outcome ("Listed workspace members → 3 members") rather than reverting to "Listing members" in the history (#3663).
• Durable agent turns stalling or writing to the wrong board — chat titles wait for the conversation to commit and give up rather than holding it, and a scheduled job may only write to a board its agent already has rights on (#3661).
• Email drafted by AI receiving its own content twice — six email prompts opened with a note listing their placeholders, and that note was being filled in as well, so every email prompt sent its content, body included, twice (#3693).

Improvements

• Lookups and rollups propagate far faster — the four "which rows link to this one?" reads behind every lookup and rollup update were scanning the whole table once per hop, per column. On a 60,000-row table that is 60,000 records examined for one match, now one. Dependent formula and rollup cells are also recomputed in one batch per table instead of one row at a time (#3668).
• The invite dialog names your main team — a caption under the People list makes clear that every new member joins the workspace's main team whichever other teams you pick (#3651).
• Copera never asks a model to collect your data — every AI request now carries an explicit refusal of data collection, and requests are marked not to be stored (#3650).
• Deleted messages leave nothing behind — a deleted message loses its text, its translations and its transcription, including messages deleted before this change, and editing a message drops translations that no longer match (#3650).
• Message text can be kept out of notifications — a new Workspace → Security switch makes push notifications and mention emails say only who wrote, never what they wrote. Vault chips never show their label in a preview (#3650).
• Session recordings mask what you type — the composer, message text, Ask Copera, agent instructions and agent memory are masked out of session recording (#3650).
• Read-only access is honoured everywhere on docs — viewers and comment-only participants could overwrite a doc's content through the API, the command palette and agent runs. They cannot now, and editors who reach a doc through a team share are no longer refused (#3650).
• The command palette enforces your workspace's rules — a member without permission to create channels, boards, folders or docs could still create them from the palette. Every high-risk command is now checked before it runs (#3650).
• Office file saving is bound to the file it came from — saves are verified against the people actually editing the document, re-checked right before the file is overwritten so access revoked mid-save is refused, and charged to the file's own workspace rather than whoever's link triggered them (#3683).