AI Teammates You Can Hire, an Ask Copera That Never Loses Your Draft & Instagram in the Inbox

The biggest week yet for the AI Workforce: a hired agent is now a real teammate with its own memory, routines, skills, budget and a rulebook you write in plain language — and it can reply in a shared inbox or hand work to another agent. Alongside it, Ask Copera stops losing what you were typing, can be steered while it is running, and ends turns honestly instead of showing a red card for something you asked it to stop. Instagram messaging arrives in the Inbox, voice channels get a shared music queue, and the Inbox finally works properly on a phone.

New Features

AI Workforce — Hire an Agent That Works Like a Teammate

An agent you hire is now a member of your workspace with the same shape as a person: it remembers, it has a schedule, it has a budget, and it answers to rules you set (#3577).

• It remembers, with discretion — an agent keeps durable facts it learns while working, and it will only repeat something to people who could have reached the source themselves. Revoke someone's access to a channel and the agent stops bringing that channel into replies they can see.
• Routines — give an agent recurring work or work that fires on an event, with quiet hours, limits, and a rule for what happens when one run overlaps the next.
• Skills — an agent's abilities are declared up front, so you can see exactly what it is able to do before you hire it.
• Teams and hand-offs — an agent can pass a piece of work to another agent and circle back with the answer, and you decide which agents each one is allowed to reach.
• One budget for everyone — agents spend from the same workspace allowance people do. A run reserves what it needs and tops it up as it goes, so several agents working at once can never jointly overspend your ceiling. When the allowance runs out, the workforce pauses, the owners are told, and routines are parked rather than failing silently.
• Rules in plain language — write what an agent may and may not do in ordinary sentences. A second model reviews each risky action against them, and every decision lands in an action ledger you can read. Risk comes from the act itself, never from the wording of the request, so ordinary work is never queued for approval.
• Approvals where the decision is made — anything held for your say-so appears in the inbox you already read, not in a separate queue.
• A full profile per agent — activity, conversations, memory, skills, access, audience, routines, budget, rules and the action ledger, each on its own tab. Memories another agent wrote are read-only and say who wrote them.
• Managers can govern — team managers can now see and manage the agents they are responsible for; only retiring an agent and granting or revoking access stays admin-only.
• Hiring is reversible and safe — choose who an agent is for, activate it, duplicate an existing hire, or share it without ever exposing its instructions. A dry run refuses rather than rehearsing something it could not actually do.

AI Workforce — An Agent That Answers Your Customers

Give an agent a shared inbox and it can read a customer thread and reply in it — with its own address, its own reply history, and no risk of answering its own messages (#3577).

• It replies in the thread — the customer gets an email from your inbox and the conversation stays in one place.
• It knows when a reply is not its business — an explicit mention reaches it directly, and a message it has no business answering is passed on rather than answered anyway.
• It reaches out too — an agent can start outreach on your connected messaging channels, comment on a row, or reply in an inbox, and each of those is a separate permission you grant on purpose.

Ask Copera — Your Draft Is Never Lost

Two people lost real work this month: one composing a long message in a new chat who clicked a notification by mistake, another just switching between two conversations. Both are fixed, on every surface (#3563).

• Drafts survive everything — switching conversations, navigating away, and a full page reload. Ask Copera had no draft mechanism at all, and channels only kept theirs in memory until you refreshed.
• A message you have not sent yet is kept too — including the one you are typing into a chat that does not exist yet, which is exactly the case that cost someone half an hour.
• Your cursor comes back where you left it — so you can keep typing instead of clicking to the end first.
• Sending clears it immediately — a sent message can never come back from a stale draft.

Ask Copera — Steer a Turn While It Is Running

You no longer have to stop a turn and start again to change its direction (#3563):

• Steer — send a course correction mid-turn and Copera picks it up at the next step and says so. Several corrections arrive in the order you sent them rather than overwriting each other.
• Stop after this step — let the current step finish, then wrap up cleanly, instead of cutting the turn off mid-action.

Ask Copera — Honest Endings and Owner Control

• Stop means stop — pressing Stop while a tool was running used to produce a red "the action failed" card with an error reference. Stopping is now reported as stopping, and only a write that was already past the point of no return says an action may have completed (#3563).
• A finished answer is never thrown away — an answer that streamed perfectly could still vanish while it was being filed. The parts that decorate a message now degrade instead of taking the message with them (#3563).
• Turns survive a release — an update used to kill whatever turns were in flight, and eight of the month's reported failures landed on the exact minute of a deploy. A turn that had not yet done anything irreversible now reconnects and carries on instead of failing (#3563).
• Long turns condense instead of overflowing — past a threshold, older context is summarised for the model while your full transcript stays exactly as you can read it. A "Context left" indicator shows where you stand (#3563).
• "Act except deletes" is remembered — it now carries into every new chat like Plan only and Ask before acting do, because it still parks every destructive action for review. "Act without asking" still resets each chat on purpose (#3563).
• Everyone can use every mode by default — the act-freely modes are now a workspace setting rather than a hard admin lock, with three choices: every member, everyone except full bypass, or admins only. The AI acts as you and can only do what you could do by hand, so this was always a supervision decision rather than a permission one — and it is now yours to make (#3563).
• A mode you cannot pick explains itself — instead of greying out, the picker says whether your workspace restricts it or the agent itself is capped, because only one of those is something you can ask someone about (#3563).
• Every failure can be looked up — the reference on a failure card now always resolves to a real record, Retry is always shown (disabled with a reason when a replay would be unsafe) rather than silently disappearing, and a Report action copies a diagnostic you can paste to support (#3563).

Ask Copera — Fewer Wrong Moves

Live testing found three places where the wording of a request invited exactly the mistake that followed (#3563):

• No more duplicate tables — asked to add a column to an existing table, Copera would sometimes create a second table with the same name and put the column there, then refuse the follow-up because two tables now shared a name. A name collision now points at the table that already exists.
• No more invented email addresses — Copera had no way to look up a real address, so it guessed one and the send was refused at the last moment. It can now resolve a person you named to their actual address, which is what you see on the card.
• Cancelling a meeting works first time — it used to need two attempts because the meeting's own identifier was not the one the cancel step accepted.
• Scheduling the meeting bot accepts a Google Calendar share link — pasting one used to fail with an English error inside a Portuguese app. Short links are resolved before scheduling, and an unsupported link is explained in your language (#3563).

Meetings — A Shared Music Queue in Voice Channels

Everyone in a voice channel hears the same thing, from one shared queue (#3584):

• Search or paste a link — add a track or a whole playlist, up to fifty in the queue.
• Shared controls — play, pause, skip, play next, reorder and clear, all synchronised for everyone connected. Guests can take part too.
• It gets out of the way — music ducks automatically while someone is speaking, and you can mute it locally without stopping it for the room.
• It lives in the meeting, not on top of it — what started as a floating dock is now a proper tile in the grid and theatre layouts, with the transport controls on the tile itself.

Inbox — Instagram Messaging

Connect an Instagram professional account and handle its messages where you handle everything else (#3575):

• Two-way messaging — send and receive direct messages, with photos, videos and documents.
• Reactions, read receipts and unsends — all reflected in the thread, in the order they really happened.
• The 24-hour reply window is respected — Copera tells you when a conversation has passed it rather than letting a send fail at the provider.
• Bought one account at a time — an Instagram account is a monthly add-on you buy when you need another slot, and disconnecting tells you plainly what happens to the slot you paid for.
• Unread badges in the sidebar — connected inbox channels now show how many conversations are waiting, and clear as you read them.

Inbox and Channels on Your Phone

The Inbox was built for a wide screen and behaved like it. On a phone it is now a proper stack:

• List, then thread, then contact — one view at a time with a back control, instead of three panes fighting for the same width.
• Folders and tabs collapse into a dropdown — rather than a tab bar that runs off the edge.
• Headers stack instead of overflowing — the title, search, members and compose controls each get a row of their own, and previews shorten with an ellipsis rather than pushing the layout sideways.

Desktop — One Ask Mode, a Movable Pill and Native Recording Controls

The desktop app's assistant and recorder both got a substantial rework (#3567):

• Chat and Vision are one Ask — a single screen-aware mode instead of two you had to choose between.
• A pill you can put where you want it — the floating bar is draggable and remembers its place per display, with a context picker that now has a close button, real icons and a size that fits its contents.
• A cleaner shortcut map — every gesture is hold-to-talk, and Ask moves to its own chord. Customised shortcuts are preserved; retired combinations honestly say None instead of advertising keys that do nothing.
• AI Notes as its own desktop experience — with meeting detection built in.
• Recording controls float above your work — the control bar, camera bubble, click highlight and area picker are now real always-on-top windows rather than page elements, so they stay put over whatever you are recording.
• Pick the screen or window inside the dialog — the New Spark dialog lists your real screens and windows with thumbnails, so you answer the question once instead of being asked again by a second picker that could not even express Area.

Sparks — A Camera Track You Can Actually Move

The separate camera track was designed for and never worked end to end. It does now (#3567):

• The camera is recorded as its own layer — nothing is painted into the capture, so the bubble can be moved and resized in the editor afterwards.
• It is composed back into playback automatically — so viewers see the bubble without anyone opening the editor, positioned exactly where you left it while recording.
• It is round, because that is what you saw — the composed bubble now matches the circle on screen instead of stretching into an oval or disappearing entirely at full roundness.
• You know when a Spark is still being prepared — the record button carries a ring while processing finishes and a toast announces when it is ready to watch, and you can start the next recording throughout.
• A recording that failed says so — the editor used to sit on "Preparing this Spark" forever; it now explains what happened and offers the retry.
• The self-view is honest — a screen-only recording no longer shows you a camera bubble that is not being recorded.

Automations — Several Triggers, a Webhook Preview and Per-Role Field Access

• More than one trigger per automation — several event triggers can now start the same automation, so related paths do not need a rule each.
• See the webhook before you send it — the builder shows the exact request that will go out, with your field tokens already resolved against a real sample row, plus a copy-as-cURL action. Delivery history keeps what was sent and what came back, so a failure is diagnosable after the fact.
• Per-role field access — set a field to view, edit or hidden per role, instead of one shared list for the whole column.
• Search by row ID — the toolbar search now matches a row's ID as well as its title.
• Friendlier condition building — checkbox pickers, directional hints on operators, and member lists that populate themselves.
• Blank mappings are left alone — an unfilled column mapping is now skipped instead of clearing the cell it points at.

Notifications — Lobby Email Controls

Choose whether lobby activity reaches you by email, with the preference surviving a temporary mute rather than being reset by it (#3574).

Website — A Support Page, and Links That Actually Work

• A real support page — copera.ai/support now exists, with the support form embedded, and cards for Copera University, the community forum and the pricing FAQ. It is available in all six site languages (#3583).
• Email footer links go somewhere — all three links in the footer of every verification, invite and password-reset email were dead. They are fixed (#3583).
• Signing up shows the current policy — the signup form and workspace wizard opened privacy and terms documents from February 2024 that had drifted from the published text. They now link to the live pages, and the old files are gone (#3583).

Developer — Run Ask Copera From the Public API

Start an Ask Copera run without the app, follow its events, approve a pending action and cancel it — five endpoints over the same engine the app uses, so there is no second behaviour to learn (#3563). It needs a new AI permission you grant explicitly to a personal token, and a headless run may only plan or ask for approval — it can never act unsupervised.

Bug Fixes

• Gantt charts losing their colours — colours were wrong after a theme change and did not repaint until you navigated away and back (#3578).
• Lookup links that went nowhere — clicking a linked record in a lookup column now opens the task it points at (#3580).
• Blank link chips — a link column with an empty lookup rendered as nothing at all instead of showing the linked record.
• Descriptions typed while creating a row — a description entered on row creation was stored as plain text and did not open as a real document. It is now seeded correctly.
• Duplicating rows in bulk failed — duplicating several rows at once produced an error instead of the copies.
• A chat composer that rewound while you typed — the saved draft could overwrite live typing in a channel composer, and an empty composer could restore a message you had already sent.
• Artifacts that would not publish — a version could be blocked from publishing by advisory style notes that were never meant to be blocking. Substantive checks still apply, and the warnings are now shown as warnings (#3582).
• Shared Spark recordings that would not download — a recording shared with you failed to download; it now downloads properly (#3579).
• Undo losing track of a column — undoing a column change could report a failure instead of recognising the column had moved on (#3581).
• Tempo colours and controls that were hard to read — the timezone and calendar controls were low-contrast in dark mode (#3582).
• Ask Copera's navigation button forgetting where you were — clicking it always landed on the start-a-new-chat screen, losing the chat or Artifact you had open. It now returns to where you were, exactly as the other sections do (#3570).
• Dialogs you could not scroll — the meeting bot request dialog clipped its schedule form with no scrollbar, and the field permissions popover pushed its buttons out of view. The permissions matrix now keeps its first column pinned and scrolls in both directions (#3571).
• Files dropped into the Inbox composer being rejected — an attachment dropped onto the message body was refused as an unsupported type before the composer could claim it (#3571).
• Generic AI action titles stuck in English — meeting AI action cards showed English fallback titles regardless of your language (#3571).
• Inbox members without an explicit role could not act — a participant with no role assigned was treated as read-only instead of as a member, unlike every other chat (#3571).
• Being rate-limited forever — a rare timing failure could leave a usage counter that never expired, permanently blocking someone from uploading. One user was locked out of Spark uploads by exactly this; the counter now heals itself (#3567).
• Crashes on drag, template preview and upload — dropping an item outside any list, previewing a template board, and an upload that answered with an empty body could each take the page down (#3572).
• A stale app version after a release — the app now reloads itself once to pick up a new version instead of failing to load part of the page (#3572).
• Automatic updates on Apple Silicon Macs — every update on macOS Apple Silicon silently refused to install: the app quit, relaunched the old version, and offered the update again. Fixed, with a build-time guard so it cannot come back (#3569).
• Admin-reported search and workspace lookups — pasted identifiers now match exactly instead of being treated as a pattern.

Improvements

• Drive listings are much faster — opening a folder now resolves its contents in a single indexed pass instead of one lookup per row, and shared and favourite views use an index rather than scanning (#3562).
• Sorting and grouping large tables — sorting by a column now uses the real cell value and an index that matches it, with language-aware ordering applied only to text columns where it actually matters. Grouping follows the same path (#3561).
• Email previews use fonts your device has — an email opened in the Inbox now renders in a system font stack instead of waiting on a webfont the preview could not load.
• Voice channel music popover holds its height — search results and queue changes no longer make the panel jump under your cursor.
• Automatic to-do roll-over is off for now — the roll-over of unfinished to-dos to the next day has been disabled while it is reworked, and its controls are hidden rather than being present and inert.