Skip to main content
Read this page before you build a workflow you rely on. It lists the behaviours that most often catch people out, grouped by where you meet them, and links each one to the page that covers it in full. For every raw number in one table, see the limits reference.

Accounts and access

Connected accounts are personal

A step that uses a connected app always works through the account of the person the run belongs to, never a teammate’s. Owners and Admins cannot use someone else’s account either. When you run a workflow someone shared with you, Your accounts for this run asks you to choose your own account for each connected step before Run with my accounts. See Run a workflow shared with you.

An admin enables each app first

Members can connect an app only after an Owner or Admin has connected it once from Integrations → Connectors. Until then, connecting shows “A workspace administrator must enable this provider before you can connect a personal account.” Inside a run, the step fails with “Personal connection authority could not be verified. No provider action was performed.” See Enabling a provider for members.

An account’s level applies to a workflow too

A step that writes to a connected app runs from the account selected for it and follows that account’s level, with any workspace ceiling applied — the same levels that govern Ask and Signals: on Act it runs, on Draft it waits on an approval card (the step reads Awaiting approval and the rest of the run carries on), and on Read it takes no action and the step fails with “Not run: the account is set to Read, so Hive only reports what it found”. On Act, what can still make a workflow’s action wait is where it sends: see When a workflow action waits for approval. New accounts start on Act, and each member sets the level of their own accounts, whatever their role. See What agents and workflows do by themselves.

Whose run it is

Every run needs the person it runs as to hold the Operator role or above and a Basic, Pro or Max seat, and a public form only starts runs while its owner still has both. A run its owner’s missing seat stops is skipped and recorded, and starts again once they have a seat. See Whose accounts a schedule uses and Runs from a form.

Runs and costs

Test runs are real model calls

A Test run runs your draft. It calls AI models and uses credits exactly like a production run (on Business, your own AI credits; on an earlier plan, the workspace’s pool, never test credits), and it counts toward the workflow’s Monthly credit limit. It does not count as one of your workflow runs. Usage lists it as Workflow tests, apart from live runs. Actions are not sent: Past runs shows Not sent: this was a test run and what the step would have sent, and an Agent library step takes none of its agent’s actions but says what it would do. A test also skips timed Wait steps and asks no one at an Approval step, carrying on as if it were approved. A test does check who each send goes to: when a production run would refuse an email address, a Slack channel or a Slack direct message, the test step fails with the reason and the test stops there (see Test runs check recipients). See Test runs.

Costs are in GBP

Each run’s detail in Past runs shows what its model and connected-app calls cost, in GBP, at 1p a credit. The Model cost on each All workflows row covers model calls only, across every run you can see since 12 October 2026, when Hive began taking payments, test and failed runs included. Runs before that date are left out of the total, though each one still shows its own cost in Past runs. A step whose model cost Hive could not work out shows Cost unknown, never £0.00. See Follow a run and Billing and plans.

Every live run counts as a workflow run

On Business, every seat includes 2,500 workflow runs a month, and the execution add-on adds more for the whole workspace. A live run, however it starts, counts once against the runs of the member it is charged to (yours for a run you start, the owner’s for a scheduled, app-event or public form run), then against the add-on. Test runs do not count, and nor does a run that resumes after a wait. Past both, a run you start is refused until next month, and an unattended run is refused and recorded (see below). See Workflow runs.

Monthly credit limits

On Business, a workflow has no monthly credit limit unless you set one: give it a Monthly credit limit on the workflow’s Settings tab, under Limits, alongside a Per-run credit limit and Runs per hour. On earlier plans, a workflow without its own limit gets the plan default: 2,000 credits a month on Starter and Team, and 10,000 on Business and Enterprise. Once its runs have used the limit, new runs do not start until the next billing month, and a run in progress stops at its next AI step. See Credit limits. A run Hive estimates at more than £5 (500 credits on an earlier plan) asks you to confirm before it starts, and is held to the figure you confirm for the whole run, after a Wait or an Approval step too. Deploy workflow shows what each run is estimated to cost before you deploy: in pounds on Business, in credits on an earlier plan.

Runs past 50 agent steps ask first

A run takes at most 50 agent steps (Agent and Agent library steps, each agent of a Parallel agents step, and each loop item that runs one) without asking. A test run, Run live now or a form you fill in that could take more asks you to confirm first. A scheduled run, a connected-app-event run or a public form submission pauses before its 51st agent step until an Owner or Admin approves it from their Inbox; declined, or undecided for 72 hours, it stops. See Runs past 50 agent steps.

Out of credits, unattended runs wait

A scheduled run or a run started by a connected app event that finds the workspace out of credits waits for up to 72 hours and starts by itself when credits return. A workflow with no AI step uses no credits, so its runs never wait. A schedule keeps one run waiting at a time: its later times are skipped while that run waits, so it runs once to catch up when credits return. A waiting run that starts and then fails is not started again. After 72 hours, the run stops waiting and will not start by itself: the workflow shows Run when credits return? in your workflow list, and Owners and Admins get a Run when credits return? notice. Choose Run when credits return in Settings → Billing → Runs waiting for credits and it starts as soon as credits are back. In a read-only or paused workspace, waiting runs keep waiting until it is active again. A public form submission is not kept waiting: it is recorded as a Failed run. See Waiting for credits.

Refused unattended runs still leave a record

A schedule, public form or connected app event that Hive refuses to start still appears in Past runs as a Failed run that names the reason. The member the run would have run as is always told with a card in their bell, along with anyone already listed under Notification recipients. A refused run uses no credits. Repeated refusals for the same reason are collapsed into one Failed run rather than one per occurrence, and each person gets at most one bell card per workflow an hour. Common reasons:
  • the workflow reached its Monthly credit limit or Runs per hour;
  • the plan does not include connected app events (they need Business or Team, and the 7-day trial does not include them; the workspace’s Owners and Admins are also told once, in the Inbox under Needs you), or, on Team, the workspace used its connected-app-event runs for the month;
  • on Business, the owner’s workflow runs and the workspace’s execution add-on are both used for the month;
  • a public form reached its 500 runs for the day or 5,000 for the month, a trial’s public runs, or found its owner (on an earlier plan, the workspace) out of credits; the person submitting it is told it was not sent;
  • the kill switch is on;
  • the workspace is read-only (its trial ended without a plan, its plan was cancelled, or a payment has been overdue for 14 days), or a payment is overdue;
  • the run owner’s connected account is missing, ambiguous or needs more permissions;
  • someone other than the schedule’s or workflow’s owner deployed a new version. For a saved schedule, the schedule owner reviews it and turns the schedule back on; for connected app events (and older schedules), the workflow owner reviews and redeploys it;
  • a connected app event was too large for a run to record: a text over 20,000 characters, a list over 500 items, or a field name over 120 characters. Hive does not trim the event, because the workflow would then act on data it never received;
  • a connected app event could not start the workflow because of a problem on Hive’s side. A passing fault, such as a brief database outage, is retried automatically instead, and runs the workflow once.
If the run owner has left the workspace or no longer holds the Operator role, Hive does not start the run and does not record a Failed run in Past runs. Schedules and connected app events are skipped, and the public form shows as unavailable. While the run owner holds no Basic, Pro or Max seat (theirs was moved to someone else, or suspended over the plan’s limit), each scheduled or event-started run is skipped and recorded, Owners and Admins get one notice, and runs start again once the owner has a seat. See When runs do not start.

A stuck step is started again at most 3 times

A step that does nothing for two hours, that a Hive restart interrupts, or that Hive loses track of while it runs is started again automatically, at most 3 times in total. If it is interrupted again after that, the step fails and the run fails, whatever On error says, so the step cannot keep using credits. The failed run is treated as Hive’s failure and refunded as usual, within the monthly refund limit. See Timeouts.

Results too large to record

A step’s result is recorded with its run, up to 20,000 characters of text, 500 items in a list and 120 characters in a field name. A step whose result is larger (an agent’s very long answer, say) fails with This step finished, but its result was too large to record. It keeps what it cost, and anything it already did is not undone or retried; the step’s error handling decides whether the run carries on. An Approval message filled in from earlier steps is cut to 20,000 characters instead, so the run still pauses for approval. An Output value filled in from earlier steps is cut to 20,000 characters too, unless it reads an Agent library result, directly or through another Output step; then the step fails instead.

Paused means no production run at all

While a workflow is paused, no production run starts from any source: not a schedule, a form, Run live now, Run with my accounts or a connected app event. Runs already in progress carry on, and test runs still work. The kill switch is different: while it is on, test runs are refused too, and a refused test run leaves no record. See Pause and resume.

Building

Graph shape

A workflow holds up to 50 cards and 100 connections, has exactly one trigger, and cannot contain a cycle. Every step must be reachable from the trigger. See How a workflow is organised.

Steps after a Loop run once per item

Every step you connect only after a Loop runs inside it, once for each item. That includes steps further down the chain, such as a summary or an Output. Inside the loop, read the current item by the loop’s Name for each item, for example {{item}} or {{item.text}}. To run a step once after the loop has finished, connect it from the end of the loop’s chain as usual, and also connect it from a step before the loop, such as the trigger. The step then waits for the loop to finish and can read {{<loop id>.items}}, count, truncated and, with Collect outputs on, iterations. A deploy is refused if a step inside a loop reads that loop’s output, and the message tells you which of the two fixes applies. See Loop.

Parallel loops

With Parallel execution on, up to 5 items run at once and finish in any order. A parallel loop cannot contain a Wait, a Parallel agents step or another parallel loop; a deploy is refused until you turn off Parallel execution or move the step out. See Loop.

Agent library steps in loops and parallel steps

One run can have at most 2 Agent library agents working at once and start them at most 20 times an hour. The Agent library agents in one run can make at most 40 model calls between them, and each uses at least one. A deploy is refused when the workflow could need more than that in a single run:
  • a Parallel agents step with more than 2 library agents;
  • a loop with Parallel execution on that could run a library agent on more than 2 items at once, including a library agent inside another loop within it (the parallel loop runs up to 5 items at once, never more than its Max iterations);
  • Agent library steps and loops that could start library agents more than 20 times in the run. Every loop multiplies the steps inside it by its Max iterations.
Steps on separate branches run one after another, so they do not add up at once. A turned-off step is not counted. The message names the step, and for a parallel loop names that loop, with the fix: use fewer Agent library agents in the step, use fewer Agent library steps, lower Max iterations, turn off Parallel execution on the named loop, or split the work across workflows. Separately, the same library agent can run at most 4 times at once across all your runs. Deploy cannot check this one. If a run reaches an Agent library step while that agent is already running 4 times, the step is refused and the run stops, even with Continue with empty output set; nothing waits for a free slot. See Agent library and Limits reference.

Agent library steps after waits and approvals

Each Agent library step gets its own run time: its agent’s time budget, at most 30 minutes, counted from when that step starts. A Wait or Approval before it uses none of it, so an Agent library step after an approval decided hours later runs as normal. A run’s Agent library limits are fixed when its first Agent library step starts. If Hive’s limits for Agent library steps change while a workflow is running, its later Agent library steps keep the limits the run started with. An Agent library step that starts more than 30 days, plus the run’s Agent library time limit, after the run’s first Agent library step started is refused. All time since that first step counts, running or waiting; time before it does not.

Web search on Automatic

On Automatic, a Web search step builds its query from the text output of the steps directly before it, joined without field names and cut to 1,000 characters. It is not rewritten by a model. If the step before it is a form trigger, every field’s value goes into the query, email addresses included. To control the query, put a step that writes it directly before the search, for example an Agent step prompted to “Reply with only the search query, under 1,000 characters”, and connect the search only to that step. Inside a Loop, a search with no earlier step inside the loop builds its query from every value in scope instead, the current item included. If nothing it reads has any text, the step fails with “This web search step had no run data to build a query from.” An automatic query never carries what steps that read your connected apps produced: a Hive agent step with Read-only data sources, an Agent library step, a Read from an app step, and anything built from their output are left out, because the query goes to the search provider. When they were all the step had, it fails with “This web search step’s only run data came from steps that read your connected apps, which an automatic search leaves out. Give it a fixed query, or turn on web search in that agent step.”, and deploying warns about it first. An automatic Generate image prompt leaves them out the same way. A Fixed query is searched exactly as written, so it cannot contain {{ }} placeholders. See Web search and Generate image.

Length limits apply after placeholders are filled

A message or note body is limited to 4,000 characters both as you type it and once its placeholders are filled in. If a filled-in body is too long or fills in to nothing, the step fails with Action not permitted and the run stops, whatever On error says. The error reads “Workflow contains an action that is not allowed for this role”, so when you see it on a step that inserts earlier output, check how long that output can get. The same error appears when filled-in text carries a control character the field does not allow (a line break in a single-line field such as an email subject, for example) or makes an email subject longer than 200 characters. A Create PDF body can grow to 200,000 characters once filled in. Beyond that the step fails with “Document too long”, and On error applies. A body whose tables hold more than 100,000 cells in total (each table counted as its rows times its widest row, and every row as at least 8 cells), or that would run past 200 pages, fails with “PDF too large”. See Action and Create PDF.

Internal notification stays inside Hive

An Internal notification sends nothing outside Hive, not even to the channel you name. It records an entry in the audit log, and the channel is only a label on that record. To tell someone, use an Integrations step that posts to a connected app. See Action.

PDFs use standard fonts

Create PDF prints any character outside the Western European (Windows-1252) set as ?, such as Polish ł, Greek, Arabic, Hebrew, Chinese or Hindi letters. The euro sign €, ™, Œ, Š and Ž are kept. Smart quotes, en and em dashes, the ellipsis, bullets and the symbols →, ≠, ≤ and ≥ are swapped for plain equivalents rather than dropped. For other scripts, make the document in Ask: its PDF download keeps every script. See Create PDF.

Agent steps only look up connected data

An Agent step reads its Read-only data sources itself, as often as it needs, for up to Max Steps rounds of reads and web searches, and can read only the accounts set on the step. If an app returns no usable data, the step fails with “The selected connector account returned no usable data.” The agent cannot write to those apps; to change something in an app, add an Integrations step after it. See Hive agent.

Templates and import

Templates open before their apps are connected

A template labelled Connect and an app name can still be used: Use template opens it as a private draft, and the steps that use that app have no account yet. The Set up guide’s button for each app, such as Connect Outlook, connects it without leaving the guide; if you skip it, Finish later tells you how many apps are left. Until every step has an account, test runs are refused, and a deploy is refused with “Connect Outlook to continue.” (with the app’s name). A Not available yet template’s button reads Open as draft: it opens as a draft, without the Set up guide, but you cannot deploy or run it. See Readiness labels.

Import keeps only your own account choices

Import workflow keeps a step’s account choice only when it names one of your own connected accounts, so re-importing a workflow you exported keeps the accounts you chose. Every other account choice in the file, such as a teammate’s account, is removed. For those steps, if you have exactly one account for the app, Hive selects it for you; otherwise the step needs an account. The import message reads “imported as a draft. Fix before deploying:” and names the first step’s problem: “Choose which Outlook account to use.” if you have more than one account for the app, or “Connect Outlook to continue.” if you have none. The setup checklist under the builder’s next step shows the steps still waiting. See Copy, export and import.

Actions and recipients

When a workflow action waits for approval

In a production run, an action step is carried out on the account selected for it and follows that account’s level: on Read it is not run, and on Draft it waits as an approval card. On Act it runs, unless where it sends came from somewhere the workflow does not trust to choose it: a Public form (apart from an acknowledgement email to the visitor), a Connected app event, a record a Read from an app step read, an AI step’s answer, a Web search step, or a run an AI client started. Then the action waits as an approval card instead, in Review decision on the Signals page of the person the run runs as. An Approval step earlier in the workflow does not vouch for it. Otherwise a run waits for a person only at an Approval step its author added. See When an action waits for approval and What agents and workflows do by themselves. A held step does not fail and does not pause the run, so a loop goes on to its next item. Each held item makes its own card: a loop that sends to the sender of each of 30 inbox records raises 30 cards. There is no limit on cards per run; a run is bounded by its 500 steps. Deploying warns on such a step: “This step will wait for approval once for each item: where it sends comes from a record a Read from an app step read.” Workflows already deployed follow this from their next run, with no need to deploy them again. Your workspace’s safety rails still apply: the kill switch, the recipient checks below, the hourly action limit and the spend caps. An action one of them stops is not sent: its step fails with the reason, is not retried, and the run fails, even when the step is set to continue on error. See Why an action was not sent. A test run sends nothing, but a send that the recipient checks would refuse fails the test step too.

Who an email step may send to

A Gmail or Outlook send, an Outlook forward or a DocuSign envelope step can send to any valid email address: one written into the step, taken from the trigger, read by an Agent step or written by one. An address that is not a single valid email address is refused, in a production run and in a test run: the step fails with the error Shadowed: recipient not on record, not dispatched. In a test run the step’s detail starts Test run: not sent. (see Test runs check recipients). A Slack direct message still goes only to a workspace member, and a channel post only to a channel your connected Slack reaches (see Slack channels and direct messages). Where a valid address came from decides whether the step sends by itself or waits for approval. See Action and Know who Hive may email.
The recipient check accepts any well-formed email address, wherever it came from. On a Public form, a form field holds whatever the visitor typed, and a Connected app event carries whatever the outside sender put in it. That text can also steer an Agent step or an AI condition. Hive holds a send to an address from those places for your approval, but two kinds of send still go out with no card:
  • A public form’s acknowledgement. An email to the address a visitor typed sends from your mailbox when everything else in it is text typed into the step or values from that same form. Anyone with the form’s link can use it to send their own words to any address they type, up to the form’s submission limits.
  • A Hive agent step that only reads your apps. The answer of a Hive agent step with Read-only data sources and no Tools counts as the workflow’s own, even when the app it reads holds text strangers write, such as an inbox or a helpdesk.
To protect who receives it:
  • write the address into the step wherever you can;
  • for a form, also limit who can submit: use an In-app form, which members fill in from Hive, rather than a Public form, and share a public form’s link only with people you trust. Reset link ends the current link (see Find or reset the link).

Slack channels and direct messages

A Slack step may post only to a channel the chosen Slack workspace reaches, and send a direct message only to a member of that workspace. Hive checks each step against the Slack account the step runs from, and a direct message goes to that member’s account in that same workspace. To check a channel post, Hive asks Slack for the channels that account can reach. A direct message to a member of your Hive workspace is accepted by their email address; for anyone else, Hive asks Slack for the members the account can reach. A refused post or message says why, so you can tell a channel the connection cannot see from a connection that is missing a permission, can no longer be used, or could not be checked because Slack was busy. If Slack lists the channels but Hive cannot read the members, a channel post still goes ahead, and only direct messages to people outside your Hive workspace are refused. The same holds for Slack messages Ask and meeting recaps send. A direct message to a Hive member whom Slack has no member for fails when it is sent, and a test run does not catch it. Each message and its fix is in Slack posts and messages Hive refused. If you have connected more than one Slack account, choose the account in the step’s Connection setting. A Slack step with no account chosen is not checked against any of your workspaces, so its channel post or direct message is refused with Nothing was sent: more than one Slack connection could send this., even when the channel exists in one of them. When the chosen account has been disconnected or is no longer active, the step is refused with Nothing was sent: the Slack connection this step uses isn’t available.: reconnect it or choose another, then deploy again. With only one Slack account, Hive uses it for you. A test run checks the Slack account the test uses, which is your own. A test of a workflow someone else owns can therefore pass or fail differently from its production runs, which use the workflow owner’s account. A failed test step says when it used your account.

Limits reference

Every limit in Hive, in one table.

Triggers and steps

Every step’s settings in full.

Runs and errors

Statuses, retries and the 500-step budget.

Accounts and autonomy

Account levels and enabling apps.