Blast radius: how much could go wrong
Every action has a blast-radius tier describing how wide, costly or hard to reverse it is:
The review panel turns the tier into a risk word: Low risk (B0), Medium risk (B1–B2), High risk (B3–B4).
Reversibility
Hive also classifies every action by whether it can be undone within minutes (the “five-minute undo test”):- Reversible — drafts, internal notes, document edits.
- Compensatable — can be followed by a correcting action: sending an email or message, pausing or resuming, enrolling, escalating.
- Irreversible — refunds, charges, revoking access, deletions. Any action type Hive does not recognise is treated as irreversible.
The autonomy ladder
Each action type sits on one rung:
Moving one action type up the ladder never promotes any other. In Ask and Signals, an account set to Act
treats a Suggest or Approve each action as Auto in limits, and a Shadow action stays
observe-only. To keep an account’s actions on approval, set it to Draft.
Autonomy is earned one action type at a time.
The decision gate, in order
For every action Hive proposes in Ask (a chat with no agent picked) or Signals, its policy engine checks, in this order:1
Kill switch
If the workspace kill switch is on, the action is shadowed — recorded, never run. Nothing gets past this.
2
Account level
The level of the account the action would use. Read → Hive
tells you what it found and takes no action. Draft → nothing runs without a person’s approval.
Act → a Suggest or Approve each action is treated as Auto in limits, so it runs on its
own only if every check below passes. Bulk actions, saving a file to Drive, tools on MCP servers
and a plan you ask for with Act on this in Signals still come back as an approval card.
3
Rung
Shadow → shadowed. Suggest or Approve each → sent to a person for approval.
4
Limits (Auto in limits and Autonomous)
The action may run on its own only if all of these hold: the amount is within the per-action spend cap;
the tier is at or below the workspace’s tier threshold; confidence is at or above the confidence threshold;
the action is not irreversible; and the hourly action rate, per-action and windowed spend caps
and error budget all pass. Anything that fails goes to a person instead. If an action type’s error budget is
exhausted, it drops one rung.
5
Overrides that always win
Even when everything above passes, Hive always asks a person for B4 actions, for irreversible
actions, and for B3 actions below the high-confidence threshold.
Default limits
Hive uses these defaults. You can set your own spend caps; the other limits are fixed, and Hive can raise the confidence thresholds (and lower the tier limit to B2) for a while when too many recent automatic actions could not be verified:
The confidence and tier limits apply in Ask and Signals. An agent’s selected action or a workflow’s action step
on an Act account runs at any tier and confidence, but the spend caps and the hourly action limit still apply.
Owners and Admins can change the spend caps under Money guardrails · hard ceilings on the governance page —
see Safety and control.
Where you manage autonomy
Open Settings → Products → Governance (the Policy & autonomy page). It has these sections:- Blast radius — the tier definitions, for reference.
- Per-action autonomy · the safety ladder — the rung recorded for each action type. Changing a rung needs the Admin or Owner role. Today a change here is recorded as your policy and in the audit trail; the live gate still uses each action’s built-in rung. When the workspace has no configured action types, the section reads No actions configured yet.
- Money guardrails · hard ceilings — the Per-action cap and Daily cap.
- Integration accounts · Read, Draft or Act ceilings — Owners and Admins can cap a provider’s accounts at Draft or Read — in Ask, in Signals, and for agents’ and workflows’ selected actions; a ceiling only ever lowers a member’s own choice. See Workspace ceilings.
Workspaces taking part in the Signals pilot cannot raise an action type’s rung while the pilot runs: promotions
are frozen, and lowering a rung is always allowed.
Routine signals on an Act account
When the account a signal came from is set to Act (see Read, Draft or Act), Signals handles routine signals on that account by itself, as the account’s owner. A signal is routine when its rule asks for review only because it grants no autonomy of its own, and its blast radius is known and low. Every signal on a matter is routine, whatever amount or severity it shows: the account’s setting decides, and the limits below are the guard. A chase is the exception: it always waits for you (below). Everything in the decision gate still applies: the kill switch, the spend, rate and error limits, and the overrides for high-impact, irreversible and low-confidence actions. When any of them says no, you get a card instead. The decision is recorded with the reason “Act account: routine hold released to run within limits”, and on a matter the timeline says what Hive did and why. These always wait for you, whatever the account’s setting:- an action with a high or unknown blast radius, or one that moves money, such as a refund;
- anything you asked for yourself: Act on this, a request in Ask, a meeting recap;
- a signal with an active exception rule;
- a signal read from several members’ accounts;
- a chase — a requirement chase or an Invoice to chase — as a reply in your thread or as a new email;
- a reply that mentions something Hive could not verify (see Replies are checked against what Hive read).
Email replies on an Act account
A sent email cannot be undone, so every email Hive writes in Ask or Signals is blast-radius B4 and comes to you as a High risk card, on Act as on Draft. There is one exception. On a Gmail or Outlook account set to Act, Signals sends a reply inside your own email thread by itself: a reply you owe. A chase never goes out by itself, in your thread or not — neither a requirement chase nor an Invoice to chase: Hive prepares it as a draft that waits for your approval. A reply you owe goes out by itself only when all of these hold:- the thread is in your own mail on that account, and you wrote in it or were a direct recipient (never only copied);
- the reply goes to someone who already wrote in the thread or was a direct recipient of it, never to someone new or someone who was only copied;
- you have already replied to that exact address yourself, in that thread or before it began. What counts is a reply you wrote to a person’s own email that Hive still holds: a new email you started, a reply to a mailing or newsletter, or a reply to mail Hive no longer holds does not. Writing somewhere in the thread is not enough either: someone else could reply into your thread, and only an address you have written to counts. Hive counts only mail your Gmail or Outlook account itself filed as sent (Gmail’s Sent label, Outlook’s Sent Items) and goes by the provider’s own times, never by the From or Date a sender writes, so a forged email cannot pass as yours. Only mail you wrote yourself counts: a meeting response (accepting or declining an invitation), an automatic or out-of-office reply, an unsubscribe request, a read receipt or a delivery report does not, even though it sits in your Sent mail, and when Hive cannot tell, it does not count. An address a mailing gives for unsubscribing never counts. Mail Hive stored before this check existed does not count unless Hive reads it again, so an older correspondence may need one reply from you first. Anything else is a first reply: it comes to you as a card that says “first reply to this sender — please review”, because anyone can email you;
- the message being answered names exactly one address to reply to (one sender, and at most one Reply-To), counting every address in it, even one Hive does not show;
- Hive wrote the reply from that thread alone: to prepare it, it read that thread and nothing else — no other thread, no matter, no connected app, no Brain or knowledge search, no MCP tool — and every figure, link and address in the reply appears in that thread or in the signal itself. Otherwise the card says “Act account: this reply waits for you because Hive read more than this thread to write it.”;
- the thread had activity in the last 14 days;
- the signal was raised after this feature reached your workspace. Cards that were already waiting stay cards.
Earned autonomy
Hive notices when your team keeps approving the same kind of action. When the same pattern has been approved three times within 30 days, Hive offers to automate it.- Accepting creates an “Auto-run” rule in Learned behaviours. It becomes active straight away, so matching actions Hive prepares on its own for a signal run automatically within limits, unless the account the action uses is on Draft or Read; a plan you ask for with Act on this still waits for you. Engaging the kill switch demotes it back to Shadow.
- An earned automation can rise at most to Auto in limits. It can never reach Autonomous on its own.
- Hive never offers to automate a B4 action — sending an email, pausing a campaign, issuing a refund, revoking access — however often you approve it. A “send email” automation accepted earlier no longer appears in the list and no longer sends by itself: each email comes to you as a card again.
- A Do not run decision, or an action that ran but could not be verified, creates an exception rule telling Hive not to auto-run that shape of action.
A more rigorous, evidence-based promotion process — minimum numbers of trials per rung and measured success rates
— is being prepared. It will only ever propose a change for a person to decide; it never grants autonomy by
itself.
Promote safely
Before raising an action type’s rung:- review enough completed examples to cover normal and unusual cases;
- confirm failures were caught and corrected;
- make sure the money and volume limits are explicit;
- prefer reversible actions; and
- keep sensitive customers and high-blast-radius actions on approval.
Related
Review and approve
Decide on what Hive proposes.
Safety and control
Kill switch, policy page and spend caps.
Audit log
Every action’s receipt.
Learned behaviours
Rules, workflows and exceptions Hive has learned.