Skip to main content
Hive is built so that people stay in control of their business. The AI proposes; a deterministic policy engine — plain rules, not a model — decides whether a proposal runs, waits for a person, or is only simulated. This page covers the controls Owners and Admins use to set those rules. All of these controls live on the Policy & autonomy page. Open Settings → Products → Governance.
The Policy & autonomy page describes itself as “Legible by design: this confidence × this blast radius × your learned threshold.” Every decision Hive takes can be explained in those three terms.

The kill switch

The kill switch stops Hive acting, company-wide, immediately.
  • Engage it from the Policy & autonomy page with Kill switch.
  • While it is on, every action is forced into shadow mode (simulated, never sent), approvals are frozen, and active learned behaviours drop back to shadow. A banner across the product reads “Kill switch is on. Every autonomous action and approval is halted company-wide”, and the sidebar shows “Kill switch active · everything halted”.
  • Resume with Resume autonomy on the banner, or Actions halted — resume on the Policy page. Only Owners and Admins see Resume autonomy; for everyone else the banner ends “nothing moves until an Owner or Admin resumes it”. Resuming does not switch the learned behaviours that dropped to shadow back on: turn each one on again with Turn on in learned behaviours when you are ready.
The kill switch is per workspace, survives restarts, and is checked first on every execution path. If Hive cannot read the switch it assumes it is on — it fails closed.
Only Owners and Admins can engage or release the kill switch, and only in the Hive app: an AI client connected through Hive’s MCP server can read whether it is on, but cannot engage or release it, whatever access it was granted. Tell your team where it is before you need it.

The autonomy ladder

Every type of action sits on a rung of the safety ladder (shown as Per-action autonomy · the safety ladder). Actions start low and move up only when you allow it. The rung decides what Hive does in Ask and Signals; an agent’s or a workflow’s selected action follows the account’s level instead of a rung (see Agents and workflows). In Ask and Signals, the level of the account an action would use adjusts its rung: on Act, where new accounts start, a Suggest or Approve each action runs as Auto in limits; on Draft, every rung above Shadow asks first; on Read, Hive tells you what it found and takes no action. A plan you ask for with Act on this on a signal still comes back as an approval card, even on Act. Even on the top rungs, a small set of overrides always sends an action in Ask or Signals to a person:
  • anything at B4 (money, legal or irreversible);
  • anything Hive classifies as irreversible;
  • B3 actions below the high-confidence threshold.
Changing a rung requires an Admin or Owner. Today a change is recorded as your policy and in the audit trail; the live gate still uses each action’s built-in rung. See Autonomy and blast radius for the full decision logic and how autonomy is earned.

Integration account ceilings

Each member chooses Read, Draft or Act for their own connected accounts (see Accounts and autonomy). Integration accounts · Read, Draft or Act ceilings on the Policy & autonomy page lets an Owner or Admin cap every member’s accounts of one provider:
  • No cap — each owner chooses;
  • Up to Draft — always asks first;
  • Read only — never acts.
MCP servers are listed too. All MCP servers caps every member’s connection to any MCP server, custom ones included. Each server from Hive’s MCP directory that members have connected, such as Linear (MCP server), also has a ceiling of its own, which caps only connections to that server. A connection takes the stricter of the two. A custom server you add by address has no ceiling of its own: All MCP servers caps it, and the Policy page never shows its address. A ceiling only lowers a member’s own choice — it never raises anyone’s account — and each change is recorded in the feed. Levels and ceilings hold in Ask and Signals and for an agent’s or a workflow’s selected actions alike (see Agents and workflows).

Blast radius

The Blast radius section explains the five tiers Hive assigns to every action, from B0 (invisible or reversible) to B4 (money, legal or irreversible). The tier decides both how much autonomy is allowed and who may approve. See Roles and permissions.

Money guardrails

Money guardrails · hard ceilings caps what Hive may spend without a person, whatever the rung. The card shows your workspace’s real settings: Owners and Admins edit the Per-action cap and Daily cap on the same card. The per-action cap can never exceed the daily cap.

Signals AI budget

Signals’ own AI steps — preparing a plan or a draft, and reading your accounts for matters — use, on Business, the AI credits of the member whose accounts they read, within their top-up limit and the workspace’s monthly maximum on extra usage. On an earlier plan they run within each Max seat’s monthly Signals fair use, and above it draw on the pool up to each member’s Signals cap (see Signals fair use). There is also a separate daily safety ceiling for the whole workspace, which stops runaway AI work. It grows with your workspace’s Max seats and is never less than one seat’s worth. The ceiling holds across every Hive server, so members of the same workspace working at once can never spend past it together. If Hive cannot confirm the ceiling, for example because its counter or your workspace’s seat count is unavailable, it skips the AI step rather than run it uncapped, on every account setting (Read, Draft or Act). A new signal it would have prepared is taken off the feed as a failed preparation and is not notified. A detector rule raises it again once its preparation-failure cooldown has passed. A signal that was already handled keeps its card. A read of your accounts that could not be charged waits for the next hourly cycle. Agent Studio runs are not affected.

Circuit breakers

Behind the ladder, Hive runs circuit breakers that trip on their own. In Ask and Signals they have the effects below; an agent’s or a workflow’s action is refused instead (see Agents and workflows): Automatic execution also requires confidence of at least 0.70, and the action’s tier must be at or below the workspace’s tier threshold (B3 by default).

When Hive cannot read the limits

The breakers above depend on counters Hive reads each time it decides. If that read fails, for example during a brief database outage, Hive fails safe: it takes no automatic action. In Ask and Signals the action waits for approval with the reason Workspace limits could not be read (temporary). An agent’s selected action is refused with Hive could not check the workspace limits just now, so it did not act. A workflow’s action step retries, then fails with Workspace limits could not be read, so the action was not run. Before giving up, Hive retries a brief database fault a few times on its own. None of these means a limit was reached, so you do not need to raise any cap. Try again in a moment.

Agents and workflows

An agent’s or a workflow’s selected action follows the Read, Draft or Act level of the account it runs from, with any account ceiling applied: Read takes no action, and Draft leaves an approval card. On Act, selecting the action is the go-ahead: it is carried out straight away, whatever its rung, and the confidence and tier thresholds and the irreversible-action rule do not hold it back. What still stops it on Act: the kill switch, the hourly action limit, the spend caps, the recipient checks, an agent run’s write budget, and a pause on an action that has failed three times in an hour on the same account (see When an action keeps failing). A stopped action is refused with the reason. What turns it into an approval card on Act is where it sends: a recipient, channel, number or link that did not come from the agent’s instructions, the workflow’s own steps or words a member typed in Hive — one taken from an email, an AI step’s answer, a public form or an app event — waits for a person. A few actions also always wait for a member’s click even when selected, such as a LinkedIn post or a file replace, and a workflow step that changes something through a connected MCP server follows that server’s level. A public agent link never acts. See What agents and workflows do by themselves.

Know who Hive may email

An email Hive sends from a connected Gmail or Outlook mailbox, a forwarded message, or a DocuSign envelope can go to any email address your connected account can send to. That includes an address you typed, one a workflow supplies, one Hive found in something it read, and one Hive wrote itself. Each recipient must be a single, valid email address; anything else is refused before an approval is created, and again before sending. The other safeguards still apply to every send: the kill switch, the hourly action limit and the spend caps. In Ask and Signals, every email Hive writes waits for a person on a High risk card, whatever the account’s level (the one exception is a reply in your own thread on an Act account, described in Email replies on an Act account); the card names every recipient and where it came from, so check the addresses before you approve. An agent’s or a workflow’s selected send follows the account’s level (on Draft it waits for approval, on Read it is not sent); on Act it goes out when its recipient came from the agent’s instructions, the workflow’s own steps or words you typed in Hive, and a recipient found anywhere else waits for approval. A Slack direct message still goes only to a workspace member, and a channel post only to a channel your connected chat tool can reach. Automated invoice collections reminders are the one email exception: they still send only to a recipient on your workspace’s records, and otherwise stay a draft for you to send.
A trigger field holds whatever its sender supplied. On a workflow with a Public form or Connected app event trigger, that field holds whatever an outside visitor or sender typed, so a send to an address, channel, number or link from it waits for approval each time. The one exception is a public form’s email acknowledgement: an email to the address the visitor typed still goes out by itself when the rest of the email is text you typed into the step or values from that same form. See Where a workflow step sends.

Approvals by role

When a proposal needs a person, it appears in Signals if a signal raised it, or inline in the Ask chat where you asked for it — never both. Agents and workflows carry out their selected actions without a proposal on an Act account. They raise one on a Draft account (or under a Draft ceiling), when where an action sends did not come from words you trust, for the actions that always wait for a member’s click, for a workflow step that changes something through a connected MCP server not set to Read, and when your own message to an agent asks it to draft or hold its actions (see What agents and workflows do by themselves). It waits in the chat, or under Review decision on the Signals page of the member the run belongs to. To make a workflow wait for a person before any other step, add an Approval step: the run waits until one of the step’s approvers approves or rejects it under Needs you in their Inbox. Who can approve a proposal depends on the tier: Approving also needs an active seat: a member without one cannot approve anything, whatever their role. In your own Review decision panel on the Signals page, you can approve every card you hold yourself, at any tier: one raised by your own run, schedule, signal or workflow. A card routed to an Approver seat keeps the tier limit in the table. Approvals are bound to the exact parameters that were reviewed and expire after 7 days. If anything about the action changes, it needs a fresh approval. See Review and approve.

Everything leaves a receipt

Every action Hive takes — automatic or approved — is written to the Audit log with the signal, plan, decision and result. Reversible actions can be undone from there. Changes to who can use a workflow are written there too. When someone shares a workflow, adds a member to it, changes a member’s access or removes them, the log gets one Workflow sharing changed receipt for each member whose access changed. It names who made the change, the member affected, and their access before and after: no access, viewer, operator or editor. The change and its receipt are saved together, so a change that could not be logged does not happen: try it again. A change that leaves everyone’s access as it was writes no receipt. A sharing receipt is shown only to members who can see the workflow, such as its owner and the members it is shared with, and it shows the workflow’s current name. Anyone else in the workspace does not see the receipt, the workflow’s name or who was given access. If the workflow is deleted, or a member loses access to it, they no longer see its sharing receipts.

A sensible starting setup

1

Cap a provider at Draft while you watch

New accounts start on Act, so Hive, your agents and your workflows can take an action on their own within the limits on this page. To have them ask first for a few weeks, cap the provider at Up to Draft — always asks first under Integration account ceilings. The ceiling holds agents’ and workflows’ selected actions too.
2

Lower the money caps to match your business

The defaults are generous. Many teams start with a per-action cap in the hundreds of pounds.
3

Give approval rights deliberately

Keep B3 and B4 approvals with leads and Admins.
4

Know where the kill switch is

Make sure at least two people can reach it.

Autonomy and blast radius

How Hive earns autonomy.

Review and approve

The approval panel step by step.

Audit log

Receipts and undo.

Security overview

The controls behind the controls.