Changelog

Stay up to date with what’s new. Follow @auterynai for announcements.

November 2026

A failed Knowledge sync tells you once Fixed

A Knowledge sync that fails is retried, and every failed attempt used to send its own “sync failed” notification. You now get one notification, after the last retry. A sync that succeeds on a retry sends none. To fix a failed source, see Sync & Manage.

A deleted conversation shows a clear message Fixed

A deleted conversation used to show “The service may be briefly unavailable” with a Retry button that could never work. It now reads This conversation doesn’t exist any more. Pick another one, or start a new chat. and has no Retry button.

October 2026

Annual billing is back New

In Billing & usage, open the plans and choose Annual to pay for 10 months and get 12: Pro is $290 a year and Business $990. Your plan’s credits still arrive every month, not once a year. The launch price is for monthly plans only. See Pricing and Changing your plan.

A plan change inside the month applies at once Fixed

When you upgrade partway through a month, the new plan’s computer hours and storage now apply straight away. Before, they waited for the next month, so a paid plan could still show the Free allowance and charge for computer time it already includes. There is nothing to do on your side. If you were charged for computer time your plan includes, contact support.

Report a charge from the app New

Report a problem in the Console and the workspace now has a Billing or a charge choice. Add the date and amount of the charge and describe what you expected instead. Billing reports are filed as high priority and reach the team that handles billing. See Reporting a charge.

Card payments ask your bank to confirm Changed

Checkout now asks your bank to confirm a card payment with 3D Secure where the card supports it, so you may see your bank’s confirmation screen before the payment goes through. Some banks decline online card payments that skip this step. See Paying by card.

A new card pays an open invoice straight away Fixed

During the 7-day grace period after a failed renewal, adding or updating your card now pays the open invoice with it at once. See If a payment fails.

A failed renewal always starts the grace period Fixed

In some cases a failed renewal did not send the failed-payment email or start the 7-day grace period, so the paid plan carried on without a deadline. Both now happen for every failed renewal. There is nothing to do on your side.

The Billing page waits for your plan after you pay Fixed

After you pay at Checkout, the page now says Applying your plan. Plan changes stay paused until the new plan shows, which stops a second Checkout for a plan you already paid for. If the plan still shows Free after a minute, the page says your payment went through and you do not need to pay again. See Changing your plan.

Upgrading asks you to confirm first Changed

On a paid plan the Plan card reads Change plan. Choosing a higher plan shows the price, the date your billing period ends and that tax is added where it applies. Nothing is charged until you choose Upgrade and pay now.

The Billing page shows credits left Changed

The top bar and the Billing card now say how many credits you have left, and on a monthly paid plan when the monthly credits reset. They no longer show a percentage used. See Viewing your balance.

A refunded auto top-up turns auto top-up off Changed

When an auto top-up payment is refunded, auto top-up is switched off and the organization owner gets an email, so it does not buy credits again until you turn it back on. See Refunds.

A declined card no longer takes the launch price away Fixed

If your first payment was declined at Checkout, the next attempt showed the regular price, as if your team had already subscribed. The launch price now stays until you have paid for a subscription, so you can try again with another card and still get $19 for Pro or $79 for Business for the first 3 months of a monthly plan. See Launch offer.

Paying after a declined card no longer cancels your plan Fixed

After a declined payment at Checkout, the unpaid attempt expires on its own later. That expiry could be treated as the end of your subscription, so your plan could show as cancelled or move to Free while a second payment had gone through. Only the subscription you paid for now changes your plan. There is nothing to do on your side. If your plan or price looks wrong after a declined payment, contact support.

Credit packs and auto top-up are on Pro and Business Changed

Credit packs and auto top-up are now available on Pro and Business. On the Free plan, Billing shows Upgrade to Pro where Buy credits used to be, and the auto top-up setting is hidden. Credits you already bought stay in your balance and can be used as before. See Top-up credit packs.

A used-up computer-hours allowance no longer stops chat Fixed

When a plan’s computer hours were used up, every message was refused, even with credits in your balance. Your Agent now keeps answering anything that does not need the computer. A task that does need it says this month’s computer hours are used up. See Included infrastructure.

Not enough credits to start a run now says how many you have Changed

A run needs at least 50 credits to start. Below that, the message now shows how many credits you have and what to do: Upgrade on Free, Add credits or upgrade on Pro and Business. The balance in the header keeps showing your real number instead of dropping to 0, and the low-credits banner says how much is left instead of a percentage used.

Draft with AI says why it stopped Fixed

When a plan limit or an empty credit balance stopped Draft with AI, it showed “Couldn’t draft this one.” with no reason. It now opens a Plan limit reached dialog with the reason and an Upgrade button, and keeps what you typed. See Create an Agent.

Test connection always shows a result Fixed

In some browsers the Test connection popover could open empty. It now shows the result, or “No result from the test. Try again.” See Connect an integration.

Long runs no longer end early with “Execution timed out” Fixed

A long run could end with an “Execution timed out” error before it reached its own budget, with no way to continue it. Runs now go on until they reach their budget. When one does, the Agent writes a final summary and you can pick it up with Continue, as described in When a run reaches its budget. A long run can now use more of its credits before it ends. We’re sorry for the interrupted runs.

“Credits are low” now counts the credits you bought Fixed

An organization that had bought a credit pack could see “Credits are low” while thousands of credits remained, because the warning looked at the monthly credits alone. The warning now looks at everything you can still spend, monthly and purchased. It shows when 20% or less of your plan’s credits are left. Your balance and charges are unchanged. The usage meter at the top still counts your monthly credits. See Credits & Billing.

Voice panels are simpler Changed

Under Sample, Record and Upload a file are now one switch. Below the consent line, the clone panel says where both recordings go and who can use the voice. Design a voice needs only a description, not 100 characters. See Voice.

Runs waiting for approval no longer show as running Fixed

In the Command Center, a run waiting for approval counted as running in Live now and the Runs count. It now appears in Needs you with a Review button for owners and admins. See Autonomy & Approvals.

Saving an Agent can’t delete it Fixed

A save that set an Agent’s status to deleted deleted it, skipping the owner and admin check that Delete applies. It is now refused, with an error that points to Delete. See Deleting an agent.

Agent tokens run their own Agent, and only that Changed

POST /api/run with a token from an Agent-role API key used to answer 403. It now runs the Agent the token was issued for. The token reaches only that Agent’s run, status, reconnect, stream, run events and cancel calls, and only for tasks it started. Every other call answers 403, including listing tasks, answering input, memories and approvals: use a member token or the workspace for those. POST /api/run/async answers 409 when the task_id is unavailable, so give each new task a fresh UUID. See Agent token scope.

Terminals close with the run Changed

An interactive terminal an Agent opened now closes when its run ends. A run that is waiting for your answer keeps its terminal until it ends. Before, a terminal was closed only after it had sat idle for 10 minutes.

After you take control of the browser and release it, the Agent looks at the page again instead of repeating a click it had aimed before you took control.

Choose how long a question waits New

In Settings > Organization > Governance, the organization owner chooses how long an Agent’s question waits for a reply: Keep waiting, 15 minutes, 1 hour, 4 hours or 24 hours. After that, the Agent takes the answer it recommended. Where it marked none, it makes the safest assumption and tells you what it assumed, or stops and reports if a wrong guess would be costly. Plan and action approvals always wait for a person. See Governance.

Model choice is part of paid plans Changed

On the Free plan, the model list on an Agent’s page is shown but only Auto can be picked, and saving any other model is refused. An Agent that already has a model keeps it. You can always switch it to Auto, but on Free you can’t pick another model again.

Report a problem takes screenshots Changed

The Report a problem form now asks what went wrong and how much it affects you, and takes up to three screenshots, pasted or attached. It also opens from the sign-in page, where we ask for your email. On a short screen the form scrolls and Send report stays in view.

Helpers that outlast an Agent’s reply stay visible Fixed

When an Agent handed work to helpers and finished its reply first, the helpers kept working but their progress and final status stopped showing in the conversation. They now keep showing until each one ends. See Helpers.

A sandbox no longer sticks on “Injecting GitHub credentials…” Fixed

If you sent a message while an Agent’s sandbox was still starting, the run could sit on that step with nothing moving. Setup now stops after 30 seconds and shows “GitHub credential setup failed”, and a run now waits for a sandbox that is still starting instead of going on without your GitHub credentials. See GitHub.

Exporting a file with the same name no longer overwrites another chat’s file Fixed

When two chats in one team exported a file with the same name, such as report.pdf, the newer one replaced the older, so the older chat’s link opened the newer file. New files are now stored separately and each link opens its own file. A file that was already overwritten is not recovered.

September 2026

A long description from the home page survives approval Fixed

If you typed a long description into the box on auteryn.ai, signed up and waited for your team to be approved, the description could be lost, and you started from an empty page. A description of several hundred characters is now kept, and you land on the draft of your first Agent once your team is approved. A description lost before this fix is not recovered: describe the Agent again on the home page. See Sign up and set up your team.

Queuing a message while an Agent works no longer shows a false error Fixed

When you sent a message while an Agent was still working, the message was queued, but the message box put your text back and said it couldn’t queue it. Sending it again queued it a second time. The message now shows as queued, with no error, and runs when the current run ends. See While the agent works.

An action’s card stays “running” until the action finishes Fixed

A card for an action showed the green check as soon as the Agent finished writing the call, before the action ran, so a two-minute build showed a check for two minutes. The card now reads running until the result comes back. A card whose action never answers settles when the run ends.

Saving a Loop an Agent created keeps who it runs as Fixed

When you saved a Loop that an Agent had created for you (a follow-up, a goal pass, or a Loop built from a description) in the Loop form, the save erased how it was set up. The Loop then ran as the team’s owner or an admin instead of the person who set it up, stopped replying in the same conversation, and lost its daily credit limit. Saving now keeps all three. A Loop you saved this way before the fix has already lost them, so check who it runs as and its daily credit limit. See Create a Loop.

Your own machine no longer counts toward sandbox compute hours Fixed

A session that runs on your own machine was counted against your plan’s sandbox compute hours, although only sessions on Auteryn’s machines should be. It no longer counts, including sessions from before the fix. Those earlier minutes can still show in the compute summary; your plan ignores them. See Run agents on your machine.

Adding a Google Workspace service account or a Web Chat Widget now respects your plan’s limit Changed

Each plan limits how many Connections you can have for the same Integration. Adding a Google Workspace service account, or a Web Chat Widget from its setup dialog, skipped that check, so a team could go past the limit. Both now check it, and adding one over the limit is refused. Connections you already have are not touched. See Website channel.

GLM-5.3 Flash is no longer offered Removed

GLM-5.3 Flash has been removed from the model list. Agents that used it now run on Gemini and keep their other settings. To pick a different model, open the Agent and change its model.

A new wordmark on sign-in, sign-up and in our emails Changed

The sign-in page, the sign-up page and the header of Auteryn’s emails now show the wordmark as “auteryn.” with a coloured full stop.

On a phone call, your Agent can look up orders, check payments and text the caller New

An Agent with a Shopify Connection can look up an order during a call: its payment and delivery status, items, total and tracking. With a Stripe Connection as well, it finds the charges for that order and points out one that went through twice. On a number with a Twilio Connection, it can text the caller something to keep, such as a case number, from the number they called. A phone caller hears about an order only when they call from the number on it or give its email address, and only the buyer hears what was charged. On a call the Agent only reads: it can’t refund, cancel or change an order. Each lookup and text shows as a step in the call’s conversation, such as “Looked up order #9140 in Shopify”. See Voice.

Agents suggest Skills again, on paid plans Changed

After a successful run of eight or more actions in a shared conversation, an Agent may draft a Skill from what it just did, at most one a day and never from a private conversation. Drafts wait under Skills → Proposed by your agents, five at most. Activate makes one available to every Agent in your organization. Dismiss stops the Agent suggesting that workflow for a while. On the Free plan, suggestions stay off, as they have been since September 12. See Skills.

Invitations reach the people you invite Fixed

Invitations sent from Settings, one at a time or several at once, were saved, but the email never went out, so the people you invited heard nothing. We’re sorry. Invitation emails are sent again. If you invited someone before this fix, open Settings → People and resend the invitation, or copy its link and send it yourself. When we approve a new team, the approval email now arrives once, not twice. See Sign up and set up your team.

A new way in: sign up, wait for approval and join a team Changed

Sign-up has its own page. After you name your team and choose Request access, you stay signed in on a waiting page, and Check again asks whether your team is ready. If you came from a ready Agent on auteryn.ai, you go straight back to it once your team is approved. An invitation link shows the team, who invited you and your role before you do anything. Accepting signs you in to that team, with no team name to pick and no sign-out. Settings → People lists everyone and every pending invitation, and the owner can hand over the team with Make owner. See Sign up and set up your team.

Describe your first Agent, or pick a ready one, and start chatting New

A new team’s home page asks one question: What should your first agent do? Describe the work and Draft with AI drafts the Agent along with a first prompt for it. Or pick a ready Agent: they sit on four shelves, each with a preview, and Use this agent adds one in a single click. If the Agent needs an Integration your team hasn’t connected, you can connect it from the draft. Create and try it opens the Agent’s first chat in the same tab with its first prompt in the message box, ready to edit and never sent for you. Cancel stops a draft, and nothing is created until you choose Create and try it. After the first reply you can take a short tour, or skip it; Take the tour in the account menu brings it back. See Create an Agent.

A calmer, clearer look across Auteryn Changed

Every page now uses one typeface, and a status is a dot and a word, with colour kept for what needs you. Each Agent has one icon in its own colour, everywhere it appears. The sidebar keeps your recent Agents and chats, shows how many runs are live, and opens from a menu button on a phone or tablet.

Chats read like results Changed

An Agent’s reply opens with its icon. Your messages carry your picture or initials, and a label such as Guest, or the Channel a customer wrote from, appears where it matters. A finished run folds into one line that says how long it worked and how many actions it took; open it to see each step, each action and a peek at its result. Code blocks have a language label and Copy, tables have Copy table, charts and images have a caption, and a long answer is laid out as a document with an outline. Math shows as math. Above the message box you see at most one card and one line of status.

The computer follows your Agent Changed

The Computer tab switches to what the Agent is using: Browser, Console for the commands it runs, or Changes for the files it edits. It stops following while you type or take control, and you can switch following off. A rail of the run’s steps lets you replay what happened.

Type into a voice call from the message box Changed

While a call is live, the message box sends your words into the call and says so. Files and selections still go to the chat, because a call can’t carry them. The separate text box on the call bar is gone.

Generated files stay available Fixed

Images, diagrams, audio and video an Agent generated were removed after about a day, and an image an Agent saved from its computer disappeared once the computer stopped. A chat could then show This file is no longer available. New files are now kept. If an earlier file shows that message, ask the Agent to create it again.

Sign-in says what went wrong Fixed

When Google sign-in failed, the page said your email and password didn’t match. It now says sign-in didn’t go through and how to reach us. A short problem on our side no longer signs you out, a shared replay link opens without an account, and if signing out fails you’re told so. Report a problem on a signed-out page now says why it can’t file the report and offers to email it instead.

Your credit balance shows what is left Fixed

Billing & usage now shows the credits left of your plan’s allowance plus the credits you bought, and the percentage used and the Running low note follow the same figure. In overage the balance reads 0, never a negative number, and the overage has its own line. Balances of 1,000 credits or more show whole credits. See Credits & Billing.

Smaller fixes in chats and runs Fixed

Plan first works in a brand-new chat, and a Skill or plan choice made in a new chat stays in that chat. Retry on a failed upload works. Inline code no longer shows stray backticks, and your account card no longer stays on “Loading…”. Helpers working side by side, and messages you send while the Agent works, show as they happen. The Runs page no longer fails when one of its panels has no data, and a run is named by its chat’s title.

Choose who can use your machine, and give other Agents read-only access New

Adding a machine gives you a new install command. It checks the installer against a checksum and reads the install code from standard input, so the code never shows in the machine’s process list. If a check fails, the installer changes nothing and says what to fix. When an install code expires, select Get a new command. Reinstall gives a machine new credentials and keeps its name and its Agent.

Each machine now has a setting for who can use it: Only me, Me and admins or Everyone in the org. A machine added before this change starts as Everyone in the org if it runs Confined, and as Only me if it has Full account access. On a machine added before this change, the browser add-on stays off until the person who added the machine allows it again. A Loop that runs an Agent on someone else’s machine runs as the person who created the Loop.

One Agent has full access to a machine. You can give other Agents read-only access to its workspace, for example a reviewer reading the code another Agent is changing.

The runner no longer writes to your home folder. The sign-ins an Agent uses, its task files and its terminal history live in /var/lib/auteryn/state, and git ignores the links to it in your workspace. Your local network is blocked by default, and commands run with memory and CPU limits.

The Console shows each machine’s runner version, whether an update is available and the command to install it. A runner older than 0.1.25 shows Update required: Agents don’t use the machine until it’s updated, and runs follow the machine’s offline setting instead. When a run moves to a cloud sandbox because your machine was unreachable, the workspace says so and names the machine.

If the runner is already installed, update it with the command the Console shows. See Run agents on your machine.

Put your Agent on your website with the Website channel New

Add a website from Integrations → Web Chat Widget, choose the customer-facing Agent that answers, and paste one script tag. The snippet pins the widget’s exact version with its integrity hash, and the chat loads only on the domains you list.

Each website has its own page. You set the chat’s look, greeting, starter prompts and nudges, whether to ask visitors for contact details, and whether they can rate answers or attach files, with a preview of your unsaved changes. You can label the steps visitors see while the Agent works.

A website is read-only by default. You can allow abilities that change data one at a time, such as sending an email or opening a ticket, and the visitor confirms each one before it runs. A daily credit cap limits what a website spends (new websites start at 100 credits a day). An identity secret lets your server tell the chat who is signed in. A signed-in visitor’s conversations then follow them to other devices. Numbers shows the last 30 days: conversations, the share of rated answers rated helpful, and the top questions.

Widgets with a wk_pub_ key keep working. See Website channel.

Conversations continue correctly after you stop a run Fixed

After a run was stopped, a later conversation could lose track of its earlier turns and send them to the model again on every turn. Those turns cost more than they should have and could make a reply fail. Conversations now keep their history after a Stop. No action needed.

Stopped and failed runs are billed like finished ones Changed

A run you stop, or one that fails, is now charged the per-action fee for each web search and computer use it completed, as a finished run already was. A call cut off by Stop pays no fee. Model usage was already billed and is unchanged. Credits spent by a stopped or failed Goal pass now count toward that Goal’s credit budget, so the Goal can reach its budget sooner. They also count at their real cost toward your organization’s daily allowance for unattended work. See Pricing & Credits and Goals.

A run that reaches its step limit now tells you where it got to Fixed

A run that reached its limit on steps used to end as a plain failure. Now it keeps its partial answer and adds a short paragraph on what is done, what remains and what it would do next. You also get the Agent Task Failed notice. A Loop run or Channel reply that reaches its budget now sends that partial answer and paragraph, not an instruction to press Continue.

Continuing after a time-limit ending now starts from the partial answer.

20 new Agent templates, and every template rewritten New

The template gallery now has 77 Agent templates. New ones include Chief of staff, Weekly business review, Error triage engineer, Month-end close, Receivables follow-up and Candidate phone screener. Every template’s instructions now spell out how the Agent works, its guardrails and when the job is done, and many offer a Loop for work that recurs. Agents you already created from a template keep their own instructions. See Create an Agent.

74 built-in Skills, 18 of them new New

Each built-in Skill is now a step-by-step procedure that lists the mistakes to avoid. The five presentation Skills are now one, presentation-design. devops-deploy is now part of deploy-ops, and api-integration is retired. An internal Agent with no Skills of its own now also gets the phrases that tell it when each built-in Skill applies. See Skills.

Draft with AI builds the Agent it describes Changed

When you describe an Agent and choose Draft with AI, the draft now brings the Skills it chose, the Integrations it needs and suggestions for its Knowledge. Before you create it, you can edit its instructions. After you create it, Finish setup asks you to add their Connections. A draft never sets an Agent to Autopilot. Before, a drafted Agent started with no Skills and no list of Connections to add. See Create an Agent.

Customer-facing Agents never leave an approval waiting Changed

A customer-facing Agent no longer waits for an approval that nobody in a customer conversation can give. At every autonomy level it can send an email, add a calendar event and create a ticket or an issue, as long as that ability is ticked and none of your rules blocks it. Any other change it would have asked you to approve is refused. Internal Agents ask for approval as before. If you relied on approving these actions, untick the ability on the Integration. See Autonomy & Approvals.

Failed runs explain what went wrong in plain words Fixed

When a run fails, its notice and the email about it now give a short explanation instead of technical error text. When the cause is unknown, it says The run hit an unexpected error. Please try again. The run’s details and Failure insights show the same explanation, and Failure insights groups failures by it. Runs that failed before this change keep their original text.

Your Agent now gets exactly the Integration abilities you choose Changed

On GitHub, Jira, Confluence, Notion and Linear, the abilities you tick for an Agent now decide what it can do. Before, every Agent got the same fixed set of reads on these Integrations: an ability that changes something, such as Create Issue, stayed off even when you ticked it. Reads you had not ticked ran anyway. An Agent now reads only what you ticked and can use the changes you allowed. In Jira and Linear, Create Issue, Update Issue and Create Comment are no longer ticked by default, and we have unticked them on existing Agents. They never took effect before, so nothing that worked stops working; tick them where you want your Agent to use them. An Agent with none of an Integration’s abilities ticked keeps the read-only default. The Integration’s Abilities tab now uses the same names as your Agent’s settings. See Connect an Integration.

Customer-facing Agents no longer wait for approval to read Fixed

A customer-facing Agent’s reads from an Integration, such as listing calendar events or searching mail, were treated as changes, so they could wait for an approval nobody was there to give. They now count as reads and run straight away at every autonomy level. Actions that change something still follow your approval settings, and a phone caller still cannot make your Agent read your calendar or mail. To limit what a customer-facing Agent can read, untick those abilities on the Integration. See Autonomy & Approvals.

Your Agent fixes clipped slides before it shares a deck Fixed

When an exported presentation had content that did not fit on a slide, your Agent could share the link anyway. Now, the first time an export cuts something off, your Agent fixes those slides and exports again before it shares the link. If a slide still does not fit after that, the reply ends with a line naming it, so you can ask to shorten or split it. The first export still shows under Background work. See Presentations.

Audio, images and diagrams need credits to be delivered Changed

When your organization is out of credits, your Agent no longer hands over audio (including a reply read aloud), slide narration, images or diagrams it could not charge for. It tells you so and suggests adding credits. A piece of audio too long to charge in one go, such as a very long text read aloud, is refused before it is made. Add credits and ask again. See Pricing & Credits.

Comments on the Free plan are carried into the next run again Fixed

On the Free plan, a comment you left while a run was finishing could be dropped with a Background follow-ups paused notice, even on a day nothing had been spent. It is now carried into the next run. Answering an optional question your Agent left you also no longer counts against your organization’s daily allowance for unattended work.

Credit estimates for helpers use one figure per pass Changed

Before your Agent runs several helpers in parallel, or starts work that runs in passes, the credit estimate it shows now uses one figure for each pass, based on what a Deep run costs today. For four helpers, the estimate is now about 300 credits instead of about 1,200. What you are charged is unchanged: each run still costs what it actually uses. See Helpers.

An interrupted task no longer repeats its actions Fixed

When a platform restart interrupts your Agent partway through a task, the task now gets more time to finish before it stops.

If it is interrupted anyway after taking an action, such as opening an issue, sending a message or saving something to memory, it stops there. The chat says it was interrupted after making changes and was not run again, and lists what it did under Show details. That message has no Retry button. A Loop run stopped this way sends its owner the Loop failure notice.

If it had only been reading, it runs again and is told what its previous attempt already looked up. See Background & Long-Horizon Execution.

An action you approved just before an interruption now runs Fixed

If you approved an action just before a platform restart interrupted your Agent, the task could treat it as already done and skip it. It now runs. See Autonomy & Approvals.

Your Agent’s settings, all on one page New

Opening an Agent in the Console now shows a single page instead of a row of tabs. A rail on the left moves you between its instructions, Capabilities, Integrations, Knowledge, Skills and settings, and one save bar at the bottom tells you what you have changed and saves it in one go. Try it runs the Agent beside its settings, so you can send it a message and watch it answer without leaving the page. History keeps the recent versions of an Agent’s configuration, shows who changed what, and lets you put an earlier version back — and a burst of edits made in one sitting is kept as one entry rather than one per pause in your typing. See Create an Agent.

The Agent panel beside your chat now shows what is really saved Changed

Every control in the Agent panel reflects what the Agent actually holds, not what was last clicked: a change that fails to save no longer leaves a switch sitting in the wrong position, each control says who the change reaches, and the panel is read-only when you do not have permission to edit that Agent. The Capabilities it lists are the same ones, with the same names, that the Console shows.

“Web search off” now means off Fixed

An Agent could keep searching the web after you switched Web search off, because an older stored setting — one you were never shown — still counted as a yes. The switch is now the whole answer, and any leftover setting has been folded into it, so an Agent that searched when you had asked it not to will stop. Sorry if you turned it off for a reason and it kept going. See Web Search.

A long run is no longer marked failed when you reopen its chat Fixed

Reopening a conversation while a long run was still working — or reconnecting after a brief network drop — could mark that run failed while the Agent was in fact still running it. Checking on a run no longer ends it; it keeps working, and the chat shows where it has got to. See While the agent works.

A run stopped by the per-run credit limit now says so Fixed

When a run hit the per-run credit limit partway through, it could still be recorded as completed and send a “task completed” notification, so a half-finished run looked like a finished one. A run that stops that way is now recorded as stopped, and no longer reports success. See Pricing & Credits.

See what an Integration can do — before you connect it New

Clicking an Integration’s card in the Console now opens a panel that says what your agents will be able to do with it, what it asks for, and who operates it. Connect is still one click from the card. Once it is connected, the Abilities tab lists each ability in plain words, grouped by what it can read, what it can change and what it listens for, with the reason beside anything your Connection has not been granted — including the Integrations whose tab used to show nothing at all. The Usage tab now names the Agents and Loops that use it and the Connection each one goes through, instead of a count. The directory is grouped by what you are trying to do, with a Needed by your agents row for Integrations your Agents have switched on that nobody has connected yet, and finishing a connection now tells you which account it connected as and takes you to an Agent to switch it on. See Connect an Integration.

Agent runs use fewer credits Changed

Credits for agent runs have come down. Each agent action — a step through an Integration or in the sandbox — now costs 0.01 credits instead of 0.25, and context your agent has already seen in the conversation (cached context) bills at about 10% of the normal rate instead of 25%. In practice a quick chat message now costs roughly 5–40 credits — the low end once your agent’s context is cached, which it is after the first message — and a deep multi-step run roughly 50–500, depending on how many steps it takes. Your run receipt shows the new figures, and it no longer labels cached input with a fixed discount, because the rate depends on the model your agent uses. See Pricing & Credits.

Each model call in a run is counted once Fixed

When an agent streamed its reply, a single model call could be counted several times toward the run, so a run recorded more credit use than it actually had, and on the Free plan the per-run limit could stop a run earlier than it should have. Each model call is now counted once, when it completes, and a reply you stop partway through is counted once for what it used. See Pricing & Credits.

Describe it shows its progress, and says when it could not draft Fixed

When you describe an agent in plain words — Describe it on the home screen, or Draft with AI under Agents → New agent — the progress line now follows what the draft is doing, such as Researching the web… and Reading a source…, instead of sitting on its first line until the draft arrived. If no draft comes back, the page now says Couldn’t draft an agent with a Try again button, rather than showing only the closest templates as if the request had worked. Create skill from this chat shows the same step-by-step progress. See Your First Agent.

A sharper, smoother live browser you can watch full screen Changed

When you watch your agent browse, the view is now rendered at the full resolution of your display instead of being scaled up from a smaller image, so text on the page is legible rather than soft. The panel can be opened full screen, and the page the agent sees resizes to match the panel you are actually looking at — no more letterboxed bars or blur from the view being squeezed into a fixed shape. The stream also adjusts itself to your connection: on a fast link it runs at a higher frame rate, and on a slow one it steps down gently instead of stuttering or falling behind. See Computer Use & Browser Automation.

Deep that you choose is what starts a Goal Changed

When you pick Deep for a real request — on the Auto / Fast / Deep pill beside the send button, or as the agent’s default in its settings — the agent now treats it as a Goal: it checks its own result afterwards and can book a follow-up pass to finish, inside a credit budget you can see and stop. What changed is everything that used to do the same by accident. A greeting, an acknowledgement or a one-word reply never becomes a Goal, whichever mode is set; neither does a message the platform routed to Deep on its own — that run still gets the Deep model and budget, it simply books no follow-up passes; and neither does a scheduled run, a Loop, or a follow-up the agent booked for itself. Before this, saying “hi” to an agent set to Deep could start something that kept working and spending while you were away. Goals are on paid plans, and their passes count against your organisation’s daily allowance for unattended work. See Goals and Execution modes.

One place above the message box, in a fixed order Changed

Only three things now appear between the conversation and the message box, always in the same order: something waiting on you, one line of progress, and the background bar. Previously as many as six cards could stack there at once and stay after they stopped being true. The plan now lives in the conversation, at the turn that wrote it, as a single row — Plan v1 · 5 steps · ~150 credits with one status chip that moves from needing your go-ahead to approved, running, done, or superseded by a later version — and clicking it opens the whole plan in a Plan tab in the canvas, with a version switcher, rather than floating over your messages for the rest of the session. The goal card now says what is actually happening: In progress only while a run is working, Checking whether it’s done… while the result is being assessed, Next pass at 14:22 when one is booked, and Done, Budget reached or Pass limit reached when it has ended — and it keeps the credits it has spent up to date instead of reading zero all session. See While the agent works.

Your agent’s computer says what it is really doing Fixed

“Setting up your computer — this can take a minute or two” could stay on screen for the rest of a session, or appear over a computer that was already running because another conversation had just started one. Starting a new conversation could warn you that the sandbox was “being used by someone else” when the only thing holding it was the boot your own first message had triggered. The Files, Git and Terminal tabs could stay hidden after the computer was ready, and the Computer tab offered a button to start a machine that was already there. All of that now follows the real state: the status clears when the computer is ready, the tabs appear with it, the Computer tab connects to a running machine straight away and shows the start-up steps while one is still coming up, and “someone else” is said only when the holder genuinely is another conversation. A long browsing session also no longer stops collecting the page’s console messages partway through, so your agent keeps seeing the errors it is meant to report. See Sandboxes.

A message you queue, or send after Stop, is no longer lost Fixed

Typing while the agent worked and pressing Enter was supposed to queue your message as the next turn. Instead it was being rejected silently — no error, no queued message, nothing sent. It now queues and runs as intended. And a message you type after pressing Stop now reliably starts a new run rather than being treated as a steer for the run you just stopped; when a steer genuinely cannot be delivered, it stays on screen with the reason instead of disappearing. See While the agent works.

Finished conversations stop reporting on themselves Fixed

An ordinary conversation turn that had completed perfectly well could reappear on the background bar and, some minutes later, announce itself as having “stopped after 398s” — and the agent could then write you a long explanation of a failure that never happened. Chat turns no longer register as background work at all; the bar is for things that genuinely keep running after your turn ends. Where finished background work can wake a conversation that has already ended — still limited to a small set of organisations we run ourselves — it now does so only when there is actually something to read, and that turn appears as a system note, never as a message in your own words. See Background work.

Stopping a task now stops the commands it left running Fixed

When you cancelled a task that had started long-running commands on its computer, the workspace said it had stopped while the commands kept running — to their own time limit, and charging credits the whole way. Cancelling now reaches those commands, so what you are told matches what is happening and what you are billed for. See Background work.

Disconnecting a Connection clears its credentials from a running computer Fixed

When you disconnected a Connection, the credentials that had been handed to your agent’s computer were not reliably removed from computers that were already running — so an agent mid-session could still have held them. They are now cleared as soon as you disconnect. You do not need to do anything; if you disconnected something recently and want to be certain, stopping and restarting that agent’s computer clears it outright. We are sorry — a disconnect should mean disconnected. See Connecting an Integration.

Conversations are named after the first real thing you asked Changed

The conversation list filled up with rows called “Greeting and offering assistance” and “Initial Greeting and Assistance”, and a conversation could be renamed after whatever had gone wrong most recently. Names are now taken from the first substantive message, only once a run has finished successfully — and a name you set yourself is never overwritten.

A tidier conversation Changed

A handful of smaller things in the workspace. A new conversation no longer offers invented starter suggestions — it shows the agent’s name, its description and the message box — and starting from a template no longer drops a half-written prompt into your message box that then followed you into later conversations. The agent header carries one status dot instead of two, and it now says what it means when you hover it. A presentation fills the canvas instead of sitting small in the middle of it, on large screens especially. Costs read credits rather than cr. Mathematical arrows and symbols come through as symbols rather than raw markup. And the bell no longer notifies you about the conversation you are looking at.

Three ways to make your first agent, and a home that starts you off New

A brand-new organisation now opens on a home screen that greets you and offers three ways to make a first agent: describe what you want in plain words and get a draft to look over before anything is saved, start from a template, or start blank. Beside them is a one-line explanation of each part of Auteryn — Agents, Skills, Knowledge, Integrations, Loops and Contacts — and an offer to be shown around in five stops, which is only ever offered and never starts by itself. Make that first agent and you come back to the home screen with a Say hello card, a suggested first message, and a five-step setup list — an agent, a first conversation, a Connection, Knowledge, a Loop — that ticks itself off as you go and retires itself once it is done; owners and admins see the list. The first time you open a library page you have not met, a short card says what the noun is, gives a couple of examples, and goes away for good when you dismiss it. In the conversation there is now a welcome line on your first visit, a sentence explaining the wait the first time an agent’s computer is starting — and that it stays warm afterwards — and a quiet note when your first run finishes. See Quickstart.

A home screen built around what needs you Changed

Once your agents are working, the home screen leads with one line of plain status — agents healthy, what is running, today’s runs and how they went — and then with Needs you: approvals waiting on someone, runs that failed in the last day, Connections that have stopped working or need re-authorising, an agent that is still missing something, a low credit balance. Each has one action, and when there is nothing the lane says so rather than disappearing. Below it, Live now shows the runs in progress with the step they are on and how long they have been going, refreshing on its own, and keeps the last completed run there when nothing is running. Throughput now always renders its figures, showing a dash where a period is too thin instead of hiding the panel, and keeping the last good numbers with a note when a refresh fails. Your agents appear as tiles from the very first one, each with its recent activity, rather than only once you have several.

The chat workspace no longer dead-ends when you have no agents Fixed

Opening the chat workspace before you had made an agent showed “No agents are available for this organization” and nothing you could click, which is exactly the moment you most needed a way forward. That screen now explains what an agent is and offers to create your first one.

Helpers scale with your plan New

How far your agent can fan a job out now follows your plan: up to 4 helpers per fan-out on Pro, 8 on Business and 12 on Enterprise, with 1, 2 and 3 fan-outs running at once respectively. Free keeps single helpers. Whatever the width, only a couple of helpers work at the same moment and the rest queue — the group card now says so (Working with 8 helpers · 2 at a time), each waiting row says how many are ahead of it, and the fan-out confirmation names how many start at once. Nothing to switch on; the numbers are on the pricing page and explained under Helpers.

Stop on the background work bar works again Fixed

Stop and Stop all on the bar above the message box returned an error instead of stopping anything, so a command you had asked to stop carried on running — and carried on costing credits — until it hit its own time limit. Both buttons now work. If something genuinely cannot be stopped right then it says why, for example that the browser is in use, and Stop all tells you which items it could not stop rather than reporting a silent failure. See Background work.

Background work that finished no longer blocks the next one Fixed

When your agent left several background commands running and they finished on their own, the sandbox could refuse every new background command for the next couple of hours, saying five were already running when none were — and the command it suggested you stop had nothing left to stop. A command that ends on its own now gives its slot back immediately. See Background work.

Background work is charged for the time it was allowed to run Fixed

A background command could report a longer run than it was permitted, and that unchecked figure could reach both your charge and the duration shown on its row, so a row could say one thing while the credits said another. The charge is now bounded by the limit a background command may run for, on every path, and the duration you see on the row is the one you were charged for. See Background work.

A queued helper says how many are ahead of it Fixed

When your agent fanned a job out to several helpers, a helper waiting for a free slot could only say it was waiting, never how long the queue was. Its row now reads how many helpers are ahead of it and updates as the line moves. When your agent asks you to confirm a fan-out, the confirmation also says how many of those helpers actually start at once, instead of implying they all do. See Helpers.

Everything your agent left running, in one place New

When an agent starts something that keeps going while you carry on — a long command, a development server, a helper, a scheduled run, a document export — it now appears on a slim bar just above the message box: what it is, what it is doing, how long it has been going, and a way to open it, stop one thing, or stop everything at once. A command that runs past its time limit is moved into the background and carries on rather than being killed and thrown away, and your agent is told the moment background work finishes instead of re-running a command to find out. When you are not looking, a completion raises a bell entry, a toast if you are in another conversation, and a marker on the browser tab; whether it also reaches you by email follows your organisation’s notification preferences. One part is not on for everyone yet: letting finished work wake a conversation that has already ended, so the agent can read the result and tell you what happened, is limited to a small set of organisations we run ourselves while we watch it — where it runs, it spends credits while you are away, under a daily ceiling. See Background work.

Watch each helper work, and what it costs New

When your agent hands a piece of a job to a helper, the helper now appears in the conversation as a live card — what it was asked to do, the model it is using, the steps it has taken, and the credits it has spent so far — with the steps collapsing behind its summary when it finishes. A helper can work in the background while your agent carries on with something else, and on Pro and above an agent can fan a job out to several helpers at once, shown as a group card with a row and a running cost per helper. Stopping the run stops every helper it started. Before fanning out, the agent says plainly that it is about to spend roughly as many credits as running the job that many times, and it does not start if the estimate is more than your balance. A Skill can also be set to run as a helper, with the access you choose — read-only, edit, or shell — and a step budget. See Helpers.

See what your agent remembers, correct it, or switch it off New

Memory is no longer something that just happens out of sight. Quiet chips in the conversation say when something was recalled, saved or forgotten, and open a Memory panel listing what the agent holds, where each entry came from, and how full the memory is. You can edit an entry, forget it — deleted entries stay recoverable for a while and are then purged for good — or read its history and the evidence behind it. Three switches per agent control remembering across conversations, learning automatically, and keeping per-person preferences; Don’t remember this chat does it for a single conversation, where the agent still reads what it already knows and simply learns nothing new. A Me view shows your own memories across every agent in the organisation, you can ask your agent about things from your earlier conversations with it, and a voice call in the workspace uses the same memory as chat — while a phone call deliberately writes none, because we do not assume the caller is you. See Memory.

Run a Skill with /, and see which one your agent is following New

Type / in the message box to pick a Skill for that message. While a Skill is in effect, a Following skill pill names it in the run header and says whether you or the agent started it, and the activity rows name the Skill files it reads — so a Skill-driven run is not a black box. If a long conversation gets condensed, the Skill is re-attached afterwards and says so. Each Skill now has an Invocation setting deciding whether you, the agent, or both may start it, and a Skill can narrow what the agent is allowed to do while following it — never widen it, so adding one from a link at worst asks you more often than you are used to. Skills that ship scripts or reference documents place them in the agent’s sandbox, where nothing runs by itself. The agent’s Skills section also shows when each Skill was last used on that agent, so you can detach the ones that never are. See Using Skills.

A panel that tells you which Integrations your agent can actually reach New

The workspace side panel now lists every Integration the agent is bound to with its real state — connected, needs authorisation, unavailable with the reason, or changed since you approved it — and expands to show each ability and whether it reads, writes, deletes, or always confirms. Your agent is told up front which Connections are unavailable for a run, so it tells you which one to fix instead of improvising around it. Abilities that a service itself marks as needing a person always ask, at every autonomy level, with no option to remember your answer, and are refused where there is nobody to ask — a scheduled Loop, a customer conversation, a run delegated from a phone call. An Integration can also ask you for something mid-run, as a short form or a link to finish on its own site, and images and links it returns arrive as artifacts and links rather than being flattened into text. In the Console, a Custom Integration can now sign in with OAuth instead of a static key. See Integration health & prompts and Custom Integrations.

Undo a file your agent changed, and give each conversation its own branch New

Every file an agent edits directly is recorded with the version that was there before. Each file-change card now has Undo, the group header has Undo all in this run, and hovering any of your own messages offers to put the files back the way they were when you sent it, telling you first how many files across how many runs. Changes made by a command have no per-file version to restore — the card says so and offers to restore a checkpoint instead. While another conversation is using the sandbox, anything that needs it waits its turn instead of hanging, and the workspace names the conversation holding it. When a sandbox comes back from a snapshot you get a card saying exactly what came back and what did not: unsaved files, dropped variables, commands that were stopped. Agents working in a code repository can be given a branch per conversation, with Apply to main workspace and Discard branch in the Git tab, and helpers can each work on their own branch so two never overwrite each other’s edits. See Undo & branches.

Keep typing while the agent works New

The message box no longer locks during a run. Enter queues your message as the next turn; ⌘/Ctrl + Enter steers the current one. A queued message sitting behind other runs tells you how many are ahead, and a steer that arrives just as a run ends is carried into one follow-up turn and marked carried — or marked dropped, with the reason — rather than being lost. Some questions no longer stop the run at all: the card says you can answer any time and the agent carries on without you. When a run reaches its budget it takes one short final turn to tell you where it got to and what it would do next, with Continue to pick up from there. You can also condense a long conversation deliberately with Compact now, saying what must survive, and a run interrupted by a restart replays the steps it had already taken rather than performing them again, so an email is not sent twice. Scheduled runs are now spread across a few minutes rather than all firing on the same tick, and a recurring loop your agent proposed — not one you created — expires after a while instead of running indefinitely. See While the agent works and Scheduling.

Organisation-wide approval rules and a minimum approval level New

Console → Settings → Governance sets rules that apply to every agent in the organisation at once: a minimum approval level that an individual agent may be stricter than but never looser than, always-allow / must-ask / never rules written by action — where never is the one control that cannot be worked around from inside an agent’s own settings — and a final confirmation on spending money and handling credentials, on by default. Owners and admins can see the page; only the organisation owner can change it. The same page can have Auteryn call an HTTPS endpoint you run at key moments in a run, letting your own systems block a message with a reason, add context the agent must take into account, decide an approval, or ask for one more pass; that part is limited to a small set of organisations we run ourselves for now. See Organisation governance and Autonomy & approvals.

A task waiting on a background command no longer says it has finished Fixed

When your agent started a long command — a build, a test run, an export — and stepped away to wait for it, the conversation could show the task as complete while the work was still going. The task now stays open and says it is waiting on a background command, and picks up again when that command ends. See Background & Long-Horizon Execution.

What you say in a private conversation stays in it Fixed

Facts an agent picked up during a private conversation could be saved as knowledge the agent shared with everyone who uses it. Anything learned in a private conversation is now kept to you, and an agent-wide note can no longer be written from a private conversation at all. Nothing you need to do. See Private Conversations.

Deleting a memory takes effect immediately Fixed

A memory you deleted could still be recalled by the agent for up to an hour afterwards. Deletion now reaches the agent straight away, and it is recorded in the memory history. See Memory.

Approval rules on an Integration now apply to that Integration’s actions Fixed

A rule you set for a specific Integration — always allow, must ask, or never — was not being matched against that Integration’s actions, so the rule had no effect. Those rules now apply, including to Integrations reached through a Connection. The console also refuses a rule that names an Integration your agent is not connected to, instead of saving one that could never match. See Autonomy & Approvals.

Stopping a task stops its helpers too Fixed

When an agent had delegated part of the work to helpers running alongside it, pressing Stop ended the main task but could leave the helpers working. Stop now reaches every helper, and a helper that is left behind by a failure is cleaned up.

A failed Integration action no longer reads as a success Fixed

An Integration call that failed could still be shown in the conversation as a tick with a count of fields returned. A failure is now shown as a failure, with its reason.

Testing an Integration tells you whether the credential actually works Fixed

Test reported whether we could reach the service, which is not the same as whether your key, token or account still works — so an expired credential could show green. Test now asks the service itself, using a read-only check that costs you nothing, and reports three answers: verified, rejected (with the reason), or no verdict available. A result we could not confirm is never shown as verified. Test also times out instead of hanging on a slow service. See Connect an Integration.

Searching your Integrations no longer errors Fixed

Searching the Integrations list failed for any organisation that had a connected Integration with discovered actions. Search works again, and the connected-Integration list renders those actions correctly.

Your agent’s sandbox can’t be restored or deleted out from under a running task Fixed

Restoring or deleting a sandbox snapshot while another conversation’s run was using the sandbox could pull the working files away mid-task. Those actions now wait, naming the conversation that is using the sandbox, and the workspace shows the sandbox as busy instead of offering buttons that would break the run. See Sandboxes.

Agents no longer draft Skills on their own after a run Changed

After a longer successful run, an agent could draft a proposed Skill for you to review. That proposal step is now off by default: it ran on every qualifying run whether or not you wanted a Skill from it. Skills you create yourself, Skills added from a link, and every Skill already attached to your agents are unchanged. See Skills.

Plans you can actually read, and a second model checking risky actions New

Your agent now writes a real plan — a goal, an approach, numbered steps, its assumptions and any open questions — instead of a one-line checklist. Each step shows what it will do and whether it will need your approval when it runs. You can approve it, ask for changes, answer its questions, or edit the plan yourself. Approving covers the ordinary steps, so the agent doesn’t stop again for each one, while destructive and sensitive steps still ask at the moment they run.

At Autopilot and Guarded, a separate reviewer model now checks each write or destructive action against your original request before it happens, and brings it to you when it looks beyond what you asked for. It never relaxes the safety floor, and when it is unsure or unavailable it asks you rather than proceeding.

You can also set approval rules per agent — always allow, must ask, or never — by action, and answer a waiting request out loud during a live voice call. See Autonomy & Approvals.

Stopping a run now stops it Fixed

Pressing Stop cancelled the run’s display but could leave the work itself running for up to a minute and a half afterwards, still spending credits and still able to stop for approval on something you had already cancelled. Stop now halts the work itself.

Two runs on one conversation no longer lose each other’s questions Fixed

When more than one task was working on the same conversation — a background task you delegated by voice, a scheduled run, a recurring loop — a question from one could overwrite a question from another, leaving a run waiting on an answer you were never shown. Each waiting request now appears as its own card, labelled with the run it belongs to.

Plans no longer stop for a word in the text Fixed

Plan approval was decided by matching words in the plan’s text, so a piece of pure research could stop and wait because a step mentioned “production”. Whether a plan waits is now decided by what its steps will actually do.

Automated runs use your plan’s full capabilities Fixed

A message your agent received over a channel — Telegram, WhatsApp, Slack or the web widget — or a run started by a Loop could be answered with free-plan limits instead of your plan’s: a smaller budget for the run, and premium capabilities unavailable. Those runs now carry the same entitlements as a message you send the agent yourself. See Loops and Messaging Channels.

The live browser view reconnects Fixed

Watching your agent browse could get stuck on “reconnecting…” and never recover, and the New tab button in that panel did nothing when clicked. Both now work. See Computer Use & Browser Automation.

Cloned and designed voices cost less Changed

Speech in a designed voice now bills at 50 credits per 1,000 characters and a cloned voice at 60, down from 150 for both. Preset voices are unchanged at 50. See Pricing & Credits.

Designed voices are ready the moment you create them Changed

Describing how a voice should sound no longer sends that description to a separate voice-building service and waits. Your description now steers the voice directly, so a designed voice appears in your library immediately instead of processing in the background, and it can no longer fail after the fact. Any voice you had already designed keeps working. See Audio generation.

Sound effect generation has been retired Removed

Agents can no longer generate sound effects from a description. The capability depended on a single provider we no longer use, and none of the providers that replaced it offers an equivalent. Speech, multi-speaker audio and slide narration are unaffected. See Audio generation.

Multi-speaker and long-form audio complete again Fixed

Generating a multi-speaker conversation or a long script in a preset or designed voice was failing at the final step — after every segment had been produced — and returned no audio. The segments are now combined correctly whichever voice produced them. See Audio generation.

Slide narration plays in every browser Fixed

Narration added to a presentation in a preset or designed voice was saved with the wrong audio format label, so Safari refused to play it and its length was misreported. Narration now carries the format it was actually produced in. See Audio generation.

Video generation is more reliable Fixed

Extending or editing a video your Agent already generated now completes instead of failing partway through: the continuation step was silently breaking the link to the original clip. Longer, higher-resolution videos are now labelled as AI-generated like every other one, and saving a generated video to your workspace files now writes the file. See Content Generation.

Sound effects cost what they run Fixed

Sound effect generation used to bill one flat amount whatever its length. It now bills for the seconds you ask for, so a short effect costs less and a long one costs more, in line with the published rate. See Pricing & Credits.

Multi-language audio uses the language you wrote Fixed

Generating spoken audio in several languages from one script no longer assumes the script is English. It works from the language you declare, or the voice’s own language, so the translated audio comes out right. See Audio Generation.

July 2026

Simpler agent types Changed

There are now two agent types — Internal and Customer-Facing. The former Collaboration type has been folded into Internal: internal agents now have an Allow external guests toggle to invite people outside your org to specific conversations. Existing Collaboration agents were migrated to Internal with guest access enabled. See Agent Types.

Bring your own computer New

Internal agents on paid plans can now run on your own Ubuntu/Debian machine instead of the managed cloud sandbox. Install a signed runner that dials out over a secure outbound tunnel — no inbound ports — and the agent uses your environment, credentials, and network. This also unlocks residential-IP web fetch, so pages that block datacenter traffic load from your home or office IP. See Bring Your Own Computer.

Autonomy dial New

A per-agent autonomy level now controls how often your agent pauses for approval — from approving every plan and sensitive action to running fully autonomously — with plan-approval modes of always, smart, or never. Approvals are durable, so requests wait for you and the agent resumes where it left off. See Autonomy & Approvals.

Fork a conversation New

Branch any conversation into a new thread from any point to explore alternatives without losing the original. See Fork a Conversation.

Background & long-horizon execution New

Agents can run long, autonomous tasks in the background — asynchronous runs, self-scheduled follow-ups, and resume across interruptions — beyond a single synchronous reply. See Background & Long-Horizon Execution.

Featured MCP & expanded integrations New

New dedicated guides for ServiceNow, Telegram, Twilio / WhatsApp, and the Web Widget, plus a Featured MCP tier of ~29 hosted MCP Integrations (Sentry, Datadog, Shopify, Stripe, HubSpot, Postgres, Snowflake, BigQuery) with a managed credential model — alongside OAuth integrations and self-hosted Custom MCP.

June 2026

Premium workspace New

The sandbox workspace now includes a full file explorer with file previews, a Git Source Control panel for staging and committing changes, and a multi-tab terminal — a complete IDE-style surface alongside the chat.

Real-time voice New

Talk to your agent. Real-time voice conversations are available in the workspace on Pro plans and above, with selectable agent voices.

April – May 2026

Presentation builder upgrades

Generate polished, themed slide decks, add narration and voiceover, and publish decks to a shareable link.

More featured integrations

Expanded the catalog of one-click MCP integrations and added new agent templates and skills to the library.

Smarter new-chat suggestions

Agent-aware starter prompts now appear when you open a new conversation.

March 2026

Computer Use GA New

Computer use is now generally available on Pro plans and above. Your agent can browse the web, interact with UIs, and extract data from any website.

Custom Skills New

Create your own reusable skills and share them across agents.

Knowledge Sync Performance

Sync times reduced by 60% for large Confluence spaces and Google Drive folders.

Fix: WhatsApp template message handling

Resolved an issue where approved WhatsApp message templates weren’t being sent correctly in some regions.

February 2026

Loop Scheduling New

Schedule your agent to run on a timetable with full cron expression support. Available on all plans.

Google Drive Integration New

Sync Google Docs and Drive folders directly into knowledge collections.

Datadog & Sentry Integrations New

Connect Datadog and Sentry to let your agent monitor for alerts and read error logs.

Agent Test Panel improvement

The test panel now shows real-time action details and credit usage inline.

January 2026

Collaboration Agent Type New

A new agent type designed to work alongside your team. (Consolidated into Internal agents with the Allow external guests toggle in July 2026 — see above.)

Notion Integration New

Read and write Notion pages and databases. Also usable as a knowledge source.

Fast mode speed improvement

Fast mode responses are now 35% faster on average.

Q4 2025

Web Widget — embed a chat bubble on any website with a single script tag
WhatsApp integration
Knowledge web crawler
Skill Library — Coding, Research, Customer Support, and DevOps
Multi-organization support

Q3 2025

Auteryn public launch
Core agent runtime with Integrations (GitHub, Slack, Google Workspace, Jira, Confluence)
Knowledge support (file upload, Notion, Confluence, GitHub)
Loops with webhook and schedule triggers
Telegram deployment
REST API with streaming support