The three levels
Hive’s newer engine, which these levels drive, is on for every workspace. Your policy and
autonomy settings still bound what any level can do.
Read, Draft or Act on Hive’s newer engine
Hive’s newer engine is on for every workspace. The level of the account an action would use decides what Hive does about what it finds — in Ask (a chat with no agent picked) and in Signals:- New accounts start on Act. Right after you connect an account, its panel opens with a card right above the three buttons that says what each one means. Accounts already on Draft or Act keep their level, and accounts that were already on Read stay on Read until you change them. Only accounts you connect from now on start on Act; reconnecting or replacing an account keeps the level you chose before. An account an admin set to Act before this engine reached your workspace now acts without asking first; set it to Draft if you want an approval card.
- Reading never needs approval, at any level.
- You set the level of your own accounts, whatever your role. Nobody else — an Admin included — can change it.
- Act never skips Hive’s safety checks. The kill switch, spend caps and rate limits, the action’s blast radius and Hive’s confidence, the recipient checks and the actions your workspace allows all still apply. An action that cannot be undone, or one any of these holds back, becomes an approval card instead. The one exception is a Signals reply inside your own email thread, described next. Bulk actions, saving a file to Drive and changes through MCP servers — vendor servers and Hive’s own, such as DATEV, Datasite and HMRC — always ask first in Ask, whatever the level, because they have no undo; so does a plan you ask for with Act on this on a signal (Triage signals). An agent or a workflow step its author chose such a change for follows the server’s level, as described in What agents and workflows do by themselves.
- Every email Hive writes in Ask or Signals asks first. A sent email cannot be recalled, so an email Hive writes for you — a new email, to anyone — comes to you as a High risk card, on Act as on Draft. Chat messages (a Slack or Teams post, a direct message) still follow the account’s level.
- On a Gmail or Outlook account set to Act, Signals sends a reply you owe in your own thread by itself. It does so only when the thread is yours, had activity in the last 14 days, you were not only copied on it, and the reply goes to someone who already wrote in it or was a direct recipient, and to an address you have already replied to yourself, in that thread or before it began. Hive must also have written the reply from that thread alone: it read no other thread, no connected app and nothing in your Brain to write it, and every figure, link and address in the reply appears in the thread or the signal. Otherwise the reply waits for you, with the reason “Act account: this reply waits for you because Hive read more than this thread to write it.” A chase, such as a payment chase, is never sent by itself, in your thread or not: on Act, as on Draft, Hive prepares it as a draft that waits for your approval. A first reply to someone who emailed you out of the blue always comes to you as a card. Each reply sent this way gets a receipt in your Inbox naming the email, who it went to, the account and the rule that let it go. See Email replies on an Act account. The Accounts panel of a Gmail or Outlook account says so under Act, with a link to these rules.
- At most 60 emails a day go by themselves from a Max seat’s accounts. Once the seat has sent 60 emails on its own in a day (UTC), further emails wait for you as approval cards until the next day. Each email counts once, and only when it goes by itself: one that waits with the rest of its plan for a person is not counted, and nor is a retry. When Hive detects the same signal again later that day, the email it sends then is a new one and counts too. An email no single member stands behind (a signal several members’ accounts saw) always waits for a person.
- MCP servers you connect have a level too: choose it from the server’s card in Integrations. See MCP servers.
What agents and workflows do by themselves
An agent or a workflow takes only the actions selected for it, and each one follows the level of the account it runs from — the same level, and the same workspace ceiling, that governs Ask and Signals:
On Act, selecting the action is the go-ahead, as long as where it sends came from words you
trust. When it did not, the action waits for you as an approval card instead.
- It runs straight away when you chose where it sends. The action is recorded in the audit log. In an Ask chat the agent’s answer says what it did; a run form or workflow run records it on the run.
- It waits when something it read chose where it sends. An address, channel, number or link that came from an email, a ticket, a web page, an AI step or a stranger’s form is never used by itself: you approve it first, whatever the account’s level. See Where an action sends.
- Every start counts. The same rules apply whether you ran it, a schedule started it, an app event started it (a new email, a new deal) or a signal did.
- Your “draft” wins in that run. If your own message to the agent asks for a draft or to hold the send (“draft it”, “don’t send it”, “let me review it first”), the agent writes emails and replies as drafts and leaves its other actions as cards for you to approve, for that run only.
- Connected MCP servers follow their own level. A change an agent’s author ticked on one of your MCP servers follows that server’s level: Read takes no action, Draft waits for your approval, and Act runs it (unless where it sends is held).
Some actions always wait for you, even when selected. A LinkedIn post is published in your name, so even
an agent or a workflow that selected it only drafts it: you edit the post and approve it first,
in the chat or in your Inbox. See Post to LinkedIn.
Saving a file over an existing Google Drive, Dropbox, OneDrive or SharePoint file — Replace
— waits for your approval on a card that says so, whatever the account’s level. See
Upload files from Hive, chats and email.
And when an agent creates something new that several of your accounts could carry, it leaves a
card for you to pick the account on, unless you pinned one account with
@. See
Which account something new is created in.
In Datasite, a change that deletes something, changes who can see or do what, invites someone,
creates a project or accepts a disclaimer always waits for your click, even when an agent or a
workflow selected it. See Datasite.
In DATEV, cancelling an invoice-data job, importing bookings or master data, and reversing
bookings always wait for your click too; a workflow step that makes one pauses until you
approve it. See DATEV.
An HMRC VAT return, MTD Income Tax quarterly update or final declaration is never filed by an
agent, a workflow or an account set to Act: only a workspace admin can approve one, on their own
HMRC connection, and only when the workspace allows HMRC submissions. See
HMRC.
A workflow step that changes something in another connected MCP server follows that server’s
level, as Ask does: on Read the step stops, and otherwise the run pauses until you approve
the change. See
Use a connected MCP server.- Never twice. Only a run’s first attempt acts. If Hive retries a run, for example after a restart, the retry takes no actions, because the first attempt may already have taken them.
- The workspace’s limits still apply. The kill switch, the hourly action limit, the spend caps and, for an agent, the run’s write budget can stop an action. A workflow’s action steps have no write budget of their own. Hive then refuses the action and says why; a limit never turns it into a card. An email can go to any valid address (one the agent or workflow took from something it read, such as an incoming email, waits for your approval first); a Slack direct message only to a workspace member; a text only to a number linked to one of your clients.
- Texts have their own checks, on every level. A text to a client who replied STOP, to a number Hive may not text, or past the number’s hourly or daily limit is refused, on Act as on Draft, and never becomes a card. So is a WhatsApp message outside the 24-hour window, or a WhatsApp marketing template. See Text messages. A Telegram message is refused the same way, and also when the client has never written to the firm’s bot. See Telegram bot.
- An action that keeps failing pauses on that account only. See When an action keeps failing.
- Only selected actions. An agent takes only the actions selected for its attached apps. An agent set up before actions were chosen per app takes none until you select them.
- Public agent links never act. A shared agent link takes no actions.
- AI-written text does not ping people in Slack. When an agent posts text an AI wrote, or a
workflow posts text with a
{{…}}value in it,@here,@channel,@everyoneand user-group mentions show as plain text and a link shows its address. Mentions of one person and links to a channel still work, and a ping you typed into a workflow step’s fixed text is kept.
Where an action sends
Where an action sends is any of these: an email’s recipients (To, Cc and Bcc), a calendar invitation’s attendees, a Slack or Microsoft Teams channel or direct message, a phone number, a Telegram chat, a web address, a GitHub repository, and an address in the details of an app’s or an MCP server’s action. An agent’s selected action runs straight away when where it sends appears in words you trust:- the agent’s instructions, as saved in Hive;
- what you typed in Hive for this run: your Ask message and your own earlier messages in that chat, the Run now input, or the values you typed into a run form.
What a held action looks like
A held action is an ordinary approval card. Its reason says why it waits, for example: “This waits for you: billing@example.com did not come from this agent’s instructions or from words you typed in Hive. Hive found it in something the agent read.” When Hive could not read in full where an action sends, for example because its details are too long to check, the card says “This waits for you: Hive could not read in full where it sends.”- In an Ask chat, the card waits in the chat and the answer says so.
- Anywhere else — Run now, a schedule, an app event, a signal, a workflow — the card waits in Review decision on the Signals page of the member the run belongs to: whoever ran it, owns the schedule or owns the workflow. The run’s summary ends “Waiting for your approval: 1. Hive held these actions because where they send did not come from the agent’s instructions or your own words. Find them under Review decision on your Signals page.”
- The run carries on. The rest of the run’s actions still go out, and the held card stays until you decide it, even after the run ends. A held card draws nothing from the run’s write budget, but it counts toward an agent run’s action-step limit.
- You can approve every card you hold, at any risk, from your own Review decision panel. See Review and approve.
Where a workflow step sends
Without an Approval step, a workflow carries out its action steps by itself, unless where an action sends came from a public form, an app event, an app’s records, an AI step or a web search: then it waits for approval. An action step sends by itself to:- a recipient, channel, number or link typed into the step;
- a value from a manual run or an in-app form a member of your workspace started — the fields you filled in;
- the answer of an Agent step pinned to a connected app that uses no tools, such as the Fee Reminder template’s Xero contact address, as long as nothing from outside your workspace or from an AI step was fed into that agent step.
- a Public form, which anyone with the link can fill in (with one exception, below);
- a Connected app event, which someone outside your workspace can start;
- a record a Read from an app step read, such as the sender of an email in your inbox;
- an AI step’s answer, or a Web search step;
- a run an AI client started over MCP, whose form values do not count as yours.
Public form acknowledgements still go out. An email to the address a visitor typed into your
public form sends by itself when everything else in that email is text you typed into the step
or values from the same form. If the email also includes an AI step’s answer, an app’s data, web
text or anything else, it waits for approval. A public-form value used as any other destination —
a channel, a number, a link — always waits.
- An Approval step does not vouch for a later step. A step after an Approval step still waits when where it sends came from one of the sources above.
- One card per item in a loop. A held step inside a Loop makes one card for each item and does not fail; the run carries on with the next step. A run has at most 500 steps.
- The step says why. The run shows the step as waiting for approval with the detail “Waiting for approval: where it sends came from a public form” (or an app event, a record a Read from an app step read, an AI step’s answer, a web search, a run an AI client started). The card waits in Review decision on the Signals page of the member the run belongs to.
- Deploy tells you first. When you deploy, a step that will wait carries the warning “This step will wait for approval each time it runs: where it sends comes from a public form.” (or “once for each item” inside a loop). A step that writes to a connected app gets “This step waits for approval when an address, link or number in what it writes comes from a public form.” (or an app event, a record a Read from an app step read, an AI step’s answer): it waits only on a run where such a value actually appears, and a value you typed into the step runs by itself. See Test and deploy.
- Deployed workflows follow these rules on their next run. You do not need to deploy again.
AI clients and drafts
Where an agent or a workflow sends is something you decide in Hive itself:- An AI client cannot add where something sends. Over Hive’s MCP server, an AI client cannot add an address, number, link, channel or repository to an agent or a workflow, and cannot approve a card Hive held. It can still deny or snooze one. Because the answer of an Agent step pinned to an app with no tools (or of such a row in a Parallel agents step) may decide where a workflow sends, an AI client also cannot create or import a workflow that sends to such an answer, point a step at one, move a send to another row or output field of the same step, or change that Agent step’s prompt, apps, tools or model.
- A draft Hive writes cannot add one either. The Build chat never adds a recipient, channel, number or link the agent does not already have, and never selects a new action: type it into the instructions and tick the action under Actions yourself. The workflow chat and Describe leave such a field empty for you to fill in on the canvas. In Ask, the cards that create or change an agent or a workflow and Save as agent use only the ones you typed in that chat. Your own connected mailbox addresses always count, so “email me” works everywhere.
Workspace ceilings (Owners and Admins)
On the Policy & autonomy page (Settings → Products → Governance), Integration accounts · Read, Draft or Act ceilings lists the providers connected in your workspace. Cap a provider at Up to Draft — always asks first or Read only — never acts, and every member’s accounts of that provider stay at or below it. MCP servers can be capped all together (All MCP servers) or, for a server from Hive’s directory, one at a time (see Integration account ceilings). A ceiling only lowers a member’s own choice; it never raises anyone’s. Ceilings hold everywhere an account is used: in Ask and Signals, and for an agent’s or a workflow’s selected actions (see What agents and workflows do by themselves). The panel shows “Your workspace admin allows up to Draft” next to a capped account. See Safety and control.When an action keeps failing
If the same action fails three times within an hour on the same account, Hive pauses that action on that account. For example, three failed Gmail Create draft attempts on your mailbox pause Create draft on that mailbox. Hive counts definite failures: the app refused the action, Hive could not use the account, or the action could not be sent as it was asked. An action whose outcome is unknown, such as a timeout, does not count, and neither does a stop by the kill switch or a fault on Hive’s side.- Only that account and action. The same action on another account still runs, and so does another action on the same account. A teammate’s failures never pause your account.
-
It says why and until when. The agent’s answer, or the workflow step, says for example
Gmail "Create draft" paused after 3 failures on you@example.com; resumes at 2026-10-05 12:10 UTC. Reconnect Gmail to fix it.The last sentence gives the cause when Hive knows it:- the account needs reconnecting, for example because it lacks a permission the action needs;
- the app is limiting requests;
- the app rejected the action;
- Hive could not use the account.
- Reconnecting lifts it at once. Reconnect the account in Integrations and the next run acts straight away. Otherwise the action runs again once fewer than three failures fall within the last hour.
-
MCP tools pause the same way. A tool an agent’s author ticked on one of your
MCP servers pauses after three failures within an hour on your
connection to that server, for that tool alone, for example
Linear "Create issue" paused after 3 failures on your Linear account; resumes at 2026-10-05 12:10 UTC. Reconnect Linear to fix it.Only failures that never reached the server count: the connection needs reconnecting, Hive could not use it (it is not connected, cannot be reached, or its tools changed and need a re-sync), or the tool is no longer pinned. A tool that answered with an error may have made its change, so its outcome is unknown and does not count, and neither does a fault on Hive’s side. - A half-finished action counts for a day. Some actions write in two steps, such as creating an invoice and then sending it. If the second step fails after the first has been written, that failure counts for 24 hours rather than one. This stops repeated runs from leaving a new half-finished item every hour.
Change an account’s level
1
Open the provider
In Integrations → Connectors, select a connected tile to open its Accounts panel.
2
Pick a level
Choose Read, Draft or Act for the account you want to change. On a phone, the
three levels sit on their own line under the account’s name.
Who owns a connection
Each connection is owned by the member who connected it, and only its owner sees it in Integrations → Connectors. Only the owner can disconnect their own account. Sharing a workflow or an agent never shares the owner’s connected accounts: whoever runs it uses their own.Colleagues’ connections are not shown
Every member, Owners and Admins included, sees only their own connections in Connectors. A colleague’s connection never appears there as a card, in Connected, in Needs attention or in the count at the top, and Recommended for your firm does not say how many colleagues use a tool. Search finds a tool a colleague connected only as a gallery entry with Connect, so you can add your own account. Owners and Admins see how many accounts of each app the workspace has connected in Settings → Products → Integrations, without whose they are or what they hold.Enabling a provider for members
A provider must be enabled before members can connect personal accounts. An Owner or Admin enables it by connecting it. Until then, a member who selects Connect sees: “A workspace Owner or Admin must connect this app first. Ask one of them to connect it in Integrations, then connect your own account.” The same message shows in setup’s Connect your work step, under the app it is about. Once an Owner or Admin has connected the app, the member selects Connect again to add their own account. A connection a member already has to a provider that is not enabled cannot be used either. A workflow step or agent that needs it stops before it calls the app, with “A workspace administrator must enable this app in Connectors before it can be used.” Ask an Owner or Admin to connect the provider once to enable it for the workspace.Several accounts of one provider
When you connect more than one account of a provider — for example four Gmail mailboxes — they appear under one tile. Accounts are listed in an order, primary first. An Owner or Admin can reorder them with the arrows in the panel. The order sets which account an approval’s account picker offers first. Hive never quietly switches to another account on its own. When more than one account could carry out an approved action, the approval shows an account picker.How each account is named
The accounts panel names each account by what the app says it is, so several accounts of one app are easy to tell apart, and so is a single one. For a mailbox or a Google, Microsoft 365, Calendly or Asana account that is its email address; for LinkedIn, the member’s name (or their email when LinkedIn gives no name); for Slack, the workspace; for Stripe, HubSpot and Xero, the account, portal or organisation; for GitHub, the login. Hive reads the name once when you connect the account. An app that does not say who you are shows as “App account”, numbered when there are several. An account connected before Hive could read its name picks it up when you reconnect it. An account’s name is shown only to the person who connected it.Which account a read uses
When Hive reads an app directly (in Ask, and in a workflow’s agent steps) and you have connected several accounts of it, it reads up to three of them and labels each result with its account. Name the account in your question — “check my work Gmail for the invoice” — and Hive reads only that one. It can only name accounts you own; a name that matches none of them is not read.Which account an action uses on Hive’s newer engine
In Ask and Signals, the levels of your own connected accounts for that app decide which account an action uses:- The picker starts on the account the signal came from when Hive knows it. For example, a reply to an email starts on the mailbox the email arrived in. Otherwise it starts on your primary account.
- The note under the picker follows your pick. It names the account you picked and says what approving does there, for example “Approving creates the draft in jo@work.example. Nothing is sent.” Until you pick one, it asks you to.
- Only your own accounts are offered. Hive offers connected accounts that you own and that are at the level the card is for. When you approve, Hive checks your pick again: it must be one the card offered, still yours, still connected and not set to Read. If it is not, nothing runs and the card stays so you can pick again.
- An account the evidence names is kept. An action Hive bound to the account that holds the record always uses that account, at the level it is set to.
- Hive records the account you picked with your approval decision, and the account an action ran from with what it did.
Which account something new is created in
An action that creates something new — an email draft, a new email, an event, a contact, a post, a file — has no existing record to say which account it belongs to. With several of your accounts of that app connected, you choose the account on the card:- The card shows a Send from picker listing your accounts for that app. Pick one, then Approve. (When Hive read from just one of your accounts for the request, such as the mailbox a thread came from, it uses that account, unless your message names another one.) Hive never runs it from an account on its own, even one set to Act, and approving it without a pick (from Slack, say) is refused and the card stays.
- Name the account and the picker starts on it. Write its address or name as the accounts panel shows it, as where the action runs from: “from my Outlook account jo@work.example, draft an email to sam@client.example”, “in the Acme EU portal”, or answer Hive with just the address. The card says the account was picked from your request; your approve is what chooses it. Only your current message counts, never an earlier one. An address you write to, copy in, or ask about does not count, nor does text in quotes or a one-word name such as “Marketing”.
- Pin one account with
@to choose it outright. A pinned account is used without a picker, and on Act Hive acts from it. - Naming your mailbox names the app. With Gmail and Outlook both connected, “from jo@work.example, draft …” drafts in the app that address is connected in. A recipient’s address never decides it: “to sam@gmail.com” does not mean Gmail.
- Only your own accounts are offered, never another member’s. If Hive cannot load your accounts, it prepares nothing and asks you to try again.
- In a workflow step, nothing is picked for you: choose the account in the step’s settings, or the step stops and says so.
How levels fit with the rest of Hive’s safety
Autonomy is layered. For an action in Ask or Signals to run automatically, every layer must allow it:- the account’s level (Read, Draft, Act);
- the workspace’s per-action autonomy rung, spend caps and budgets (Safety and control);
- the action’s blast radius and reversibility — money, legal and irreversible actions always ask a human;
- the kill switch, which halts everything when engaged.
Related
Review and approve
How approvals work.
Autonomy and blast radius
The safety ladder.