Follow a run
Select Past runs in the dock under the canvas. Filter by Test or Production (or hide reruns), by date or by status, then select a run to open its detail. Test run at the top of the panel tries your draft again; it sends nothing to other apps. A run you start in the builder — from the header’s Test run, the next step, a step card or a form’s Test tab — opens Past runs on that run by itself, so its result is on screen as soon as it is recorded. If the filters hide every run, the list says No runs match these filters and offers Clear filters. While the history is loading it shows placeholder rows, and if it cannot be loaded it says Couldn’t load past runs with Retry — never “No runs yet”, which is shown only for a workflow that really has no runs. Past runs stays current while it is open: a run started somewhere else — by a schedule, a webhook, a form or another browser tab — appears within about a minute of finishing. Past runs lists your own runs and your teammates’ production runs of the workflow; a teammate’s run shows its status, timing and cost but not what it received or produced. See Who sees which runs. Each run in Past runs is titled by when it started, and shows how long it took, whether it was Test or Production, its run ID (for support) and what it cost. Under that, a run can show up to three short notes:- What its actions did, once the run has finished (completed, failed, cancelled, rejected or expired) and the workflow has at least one action step. A production run reads No actions taken, 1 action taken or 3 actions taken, counting only the actions it carried out: an action that was held back or is waiting for approval is not taken. A test run sends nothing, so it reads 1 action not sent or 3 actions not sent, counting every action it reached, and shows nothing when it reached none.
- Why it did not complete, in the same words as its detail (see below), for example Slack refused ‘Post update’. The note names nobody: where the detail names a person, the row says Someone. A completed run whose condition took a path with no step on it reads Nothing ran after ‘Check amount’ (see Completed runs that did nothing).
- 2 actions were not run, on a production run that held some of its actions back (see Why an action was not sent).
- when it started and ended, how long it took, whether it was Test or Production, and a summary, for example Invoice chaser completed.;
- for a production run, the version it ran, for example Version 1.2.0, or Version unknown if that version can no longer be found;
- what the run received, under What the run received (the form answers, event or input that started it), and what it returned, under What the run returned;
- why it did not complete, in one line under its summary, for example Slack refused ‘Post update’., This run hit the workspace’s limit on actions per hour or Nobody decided ‘Approve refund’ within 7 days. A run that completed has no such line;
- when more than one step is waiting at once, Several steps are waiting, with each waiting step and what it waits for;
- one row per step, in order, with its status, detail and any error — inside a parallel step, one row per agent. Each row leads with a plain summary of what the step did, before any technical detail (see What each step did). A loop’s steps are listed under the loop’s own row, one group per item headed Item 1 of 2, Item 2 of 2 and so on (just Item 2 while the loop is still running), each step named by its title;
- what the run’s model and connected-app calls cost, in GBP: the AI credits it drew, at 1p a credit, from its member’s own credits on Business, or from the workspace’s pool on an earlier plan. While a run is still going, this adds up what its steps have cost so far, every retry included. Cost unknown means Hive could not work out what one of its AI model calls cost.
What to do about a failed run
A failed run’s detail ends with advice that follows from why it stopped:
The other reasons, including the line on a Rejected, Expired or Cancelled run, come with
no advice in the run, because the line already says what happened. Here is what to do about each:
When a run’s inputs and results have been removed, its detail says so in their place: Removed on
Oct 31 · 8:00 AM, 30 days after this run ended, or Deleted by Ada Lovelace on Oct 31 · 8:00
AM when a workspace Owner or Admin deleted them. A step whose input Hive did not record says Not
recorded in place of its input.
What each step records as its input
Runs that start on or after 30 October 2026 record what every step was given, whether the step succeeded or failed. A step still running shows its input once it finishes, and a step that failed before it was given anything shows none. Before that date, only the trigger and Action steps recorded their input, so other steps on older runs say Not recorded. Hive records step inputs for runs from that date.
A prompt is recorded as the run filled it in. Hive marks the run data inside it as data before the
AI reads it (see Run data reaches the AI as data);
those marks are not shown here.
Recording never stops a step. A recorded input is kept within these limits:
- A credential is never kept: a value under a name such as
Authorization,Cookie,password,secret,client_secret,api_keyor any name ending intoken(access_token,refreshToken) is recorded as[redacted]. - Text longer than 20,000 characters is cut there, a list keeps its first 500 entries, and a field whose name is longer than 120 characters is left out.
- An input still larger than 32 KB once those cuts are made is recorded as
{ "notRecorded": "too_large" }. - One run records up to 2 MB of step inputs. A step past that records
{ "notRecorded": "run_budget" }and runs as normal.
Link to a run
Every run has its own link. When you select a run in Past runs, the address in your browser changes to name it, for examplehttps://app.get-hive.ai/workflows/<workflow>?run=<run ID>, so
reloading the page, or going Back and Forward, reopens the same run. Select Copy link in
the run’s detail to copy that address and share it with a teammate. Opening it opens the workflow
with Past runs on that run, even when it is older than the runs Past runs has loaded. The
copied link also names the workflow’s workspace, so it opens there for someone who belongs to
several (see Links open in their own workspace).
An older link that names no workspace is looked for in your other workspaces before Workflows opens
its list.
Going Back to the workflow without a run in the address closes Past runs, and closing
Past runs or opening another workflow takes the run out of the address.
A teammate who follows your link sees the run only if they could open it in Past runs
themselves (see Who sees which runs). Otherwise Past runs says You
can’t open this run. It may belong to someone else, be a private test run, or have been removed
after 13 months., and a link to a workflow that was deleted, or that they cannot open, says This
workflow was deleted, or you can’t open it. Neither says which reason applies, so a link never
reveals whether a run or workflow exists.
Each step row also shows:
What each step did
A step’s row starts with one line in plain words, so you can read a run without opening any JSON:- an agent step shows what it answered, up to 200 characters, or its structured answer as “Label: value” pairs;
- an app step a test run held back says what it would have done: for a message, the recipient and subject, for example Would have sent email to sam@example.com: Your invoice; for other actions, a short form of what it would have sent;
- a loop says how many times it went round, for example Ran 3 times.
How often each step has run
On the canvas, a step that has run in production shows Ran 12 times beneath it. Hover to see what it counts: production runs that ran the step. A step counts once per run, however many retries or loop items it took. Test runs and skipped steps are not counted, and a step that has never run in production shows nothing. Over the API and the Hive MCP server, reading a single workflow also returns these counts asnodeRunCounts: for each step, how many production runs you can see have
run it to an outcome.
Steps that read no connected data
A Hive agent step that reads a connected app records how much data the read returned. The step still completes, because the model did answer. The record shows whether that answer had data behind it:
An empty or partial step shows a data warning on its row, as described in
Follow a run, so you can tell “nothing matched” apart from a real answer. A filter or
date range that matches nothing also comes back empty. If you expected records, check the step’s
instructions name the right records, then run it again. Hive records only these labels and a short
reason for any gap, never the data that was read. Over the API, the label is the step receipt’s
evidence.status.
When one read in a step fails (the app refuses it, is busy or is unavailable), the step does not
fail with it. The agent is told which app it could not read and why, keeps everything earlier
reads returned, and the step is recorded as Partial, never Empty, even when the reads that
worked matched no records. The step’s answer stays exactly as the agent wrote it, so nothing about
the gap reaches an email, a document or a rule-based condition that uses the answer. Instead,
Hive records which app it couldn’t read on the step (over the API, the receipt’s
evidence.unread). A later Hive agent step or AI condition that uses this step’s answer is
told about the gap alongside its instructions, and so is the first step that uses a loop’s or a
parallel step’s output when a step inside it couldn’t read an app. A read counts as read when the
agent tries the same read again on the same account and it works; a different read, or another
account of the same app, working afterwards does not hide the failure. When a
production run completes with such a step, the person it ran as and the workflow’s notification
recipients, each if they hold a Basic, Pro or Max seat, get one bell card for it (at most one an hour per workflow), saying the workflow
couldn’t read some data. The step fails only when none of its reads returned anything usable, with
the reason of the read that failed. A read your workspace’s policy refuses, or one that would go
over the step’s budget, still stops the step.
Action outcomes
An action is what an Action step does: a message sent, an email or draft, a record created or changed in a connected app, or an internal note. Other steps are not actions, even when they save something: the files a Create PDF or Generate image step saves to the Library are not counted as actions. A run’s summary, such as “Completed with 3 action outcomes”, counts every action the run took — sent or not sent — including each item’s action inside a Loop. The run’s entry in Past runs lists up to 49 of them: actions outside a loop first, then loop items in order. Every item still has its own row in the run’s detail. The list is kept however the run ends. A run that failed, was rejected at an Approval step, expired waiting for one, or was stopped by Hive after it was interrupted still lists the actions it took before it stopped, such as a Slack message it had already posted. A run stopped before its first step lists none. Starting a run through the API or the Hive MCP server answers with the same list inoutcomes. A run that paused at a Wait step answers with the actions it took
before the pause; follow it in Past runs to see how it ends.
Who sees which runs
Anyone who can open a workflow sees its production runs in Past runs, whoever ran them — including a teammate’s Run with my accounts runs and runs from a teammate’s schedule. For those runs you see the status, when it started and ended, how long it took, each step’s name and status, and what the run and each step cost. What a run received and produced stays with the person the run ran as, and with workspace Owners and Admins: its inputs, each step’s input, output and detail, its result, its summary and any error text. When you open someone else’s run, it says who can see them in their place, for example Only Priya Shah, Owners and Admins can see this run’s inputs and results. Test runs try out a draft, so each one stays with the person who ran it, and with Owners and Admins: nobody else sees it in Past runs, its run count or its cost. Owners and Admins see every run of every workflow.Run statuses
Past runs and its status filter show these labels. The code beside each is the underlying value.
A run goes back to In progress as soon as its Wait or Approval step resumes, and stays
there while the steps after it run.
Completed runs that did nothing
Completed means the run reached its end, not that every step ran. When a Condition takes a path with no step connected to it, no step runs on that path, and the steps on the paths it did not take areskipped. If nothing else is left to run, the run ends there and is still Completed.
A condition inside a Loop ends only that item’s pass: the loop’s other items still run. The
condition’s row says so, for example Took Else (…): no step runs on that path.
If the run should have done something, open the condition’s row: an AI condition records the
model’s reason. Connect a step to the path it took, or, for an AI condition, switch on Force AI to
select a path so a run with no matching condition fails instead (see
Condition).
A run can also complete without taking any action because its steps decided there was nothing to
do, for example a filter that found no overdue invoices. When the newest production run of a live
workflow completed without reaching any action, the
next step says Live — its last run took no
action, with View run to open it and see which steps ran. That can be exactly what you
wanted; the builder only tells you so you can check.
Who is told when a run expires or is rejected
When a production run ends as Expired or Rejected, Hive puts a card in the notification bell of the workflow’s owner, the person the run ran as, and each address under Notification recipients that belongs to a member of this workspace. The member who rejected the run is not told about their own decision.- Only members who hold a Basic, Pro or Max seat are notified, and only those who can open the workflow see the card. Hive sends no email.
- Test runs never notify.
- The card says the run expired or was rejected and how it started. It never quotes the approval’s message, the earlier steps’ data or the reason given for a rejection: open the run to see them.
- Each person gets at most one such card per workflow, per outcome, per clock hour (UTC). Every run is still in Past runs. An expired card and a failure card in the same hour are separate cards.
Step statuses
Each step row shows its status as a short lower-case label. A step that finished shows ascompleted.
A step you deactivated is left out of the run entirely and gets no row.
On the canvas, each step card shows its status from the latest run in words: Done, Failed,
Skipped, Cancelled, Queued, Running, Waiting. A cancelled run is shown as
Cancelled, never as a failure, and does not ask you to fix anything.
Common failure messages
Each failed attempt records why it failed, so a step that retried shows the reason on every attempt, not only the last. The run’s own message is the reason its last failed step gave.A workflow action follows the level of the account it runs from: on Read it is not run, and on
Draft it waits for approval. Without an Approval step, a workflow on an Act account carries
out its actions 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 (see Actions that wait for approval). Any
other action is carried out, or stopped with the reason. See
What agents and workflows do by themselves.
A run waits for a person at an Approval step you add. A
scheduled run, one started by a connected app event or one started from a public form also pauses
before its 51st agent step until an Owner or Admin approves it from the card in their Inbox. If
nobody decides within 72 hours, the run stops as Expired (see Runs past 50 agent steps).
Why an action was not sent
In a test run, every action is recorded but not sent: the step reads Not sent: this was a test run and shows what it Would have sent. The step succeeds, unless a production run would refuse the send because of who it goes to; then the test step fails (see Test runs check recipients). In a production run, an action that a workspace safeguard stops is a failed step. Its row keeps the reason, the step is not retried, and the run fails, even when the step is set to continue on error:
If Hive cannot read the workspace’s limits or kill switch at that moment, for example during a
brief database outage, it does not act and does not count it as a limit. It retries the step like
any other temporary fault, and if the limits still cannot be read the step fails with Workspace
limits could not be read, so the action was not run. Run it again later; you do not need to
change any limit.
Runs from before this change may show an action that did not run as a step with a dashed icon, its
reason as a warning, and 1 action was not run on the run. Older runs may also show
Shadowed (no external effect): Hive held the action back without recording a more specific
reason.
Actions that wait for approval
In a production run, an action Hive held because of where it sends (see When an action waits for approval) does not fail its step. The step iscompleted with the outcome sent for approval
(approval_required), and the run carries on: a held step does not pause its branch. Its card
waits in Review decision on the Signals page of the person the run runs as. Until someone
decides it, the step’s detail says where the address, channel, number or link came from:
Once someone decides the card, the detail says what happened:
Deciding a card does not change the run’s status, and a cancelled run is never updated. In a
Loop, each held item has its own card and its own detail. An action an earlier version of Hive
raised for approval can still be pending; its step reads Awaiting approval until someone
decides it.
Slack posts and messages Hive refused
Before a Slack step posts to a channel, Hive asks Slack which channels the step’s Slack connection can reach, and refuses a channel that is not among them. A direct message to someone Hive already knows as a member of your Hive workspace (by their email address) is accepted without that check; for anyone else, Hive asks Slack for the members the connection can reach and refuses a person who is not among them. A refused step’s detail reads Not sent: and a short reason, and its error says what happened and what to fix:
For a direct message, the messages about a missing permission, a refused list, a slow Slack and a
problem of Hive’s own say member or member list where they say channel or channels for a
post. The messages about the connection itself (switched off, unable to read its channels, more
than one connection, not available) are the same for both. When the step’s channel is not a valid Slack channel name, the message says
that channel instead of naming it.
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. A direct message to a
member of your Hive workspace is not refused here, even when Hive cannot read the Slack members or
the connection lacks a permission to read them. If Slack then has no member with that person’s
email address, the step fails when Hive sends it, not before; a test run does not catch that case.
Email steps are checked differently and keep the Shadowed: recipient not on record, not
dispatched detail.
Test runs check recipients
A test run runs the same recipient checks a production run does before sending, on the account the test uses, and still sends nothing. It cannot catch a failure that happens only while sending, such as a direct message to a Hive member whom Slack cannot find. When a production run would refuse the send, the test step fails instead of reading Not sent: this was a test run, and the test stops there, as a production run would. The step’s error is the message a production run would show, and its detail says which kind of refusal it is:- Test run: not sent. A live run on this connection would be refused now: and the reason, when the channel, person or connection is the problem. Fix what the message names, then test again.
- Test run: not sent. Hive couldn’t check this send: and the reason, when Slack did not answer or Hive could not read the list. When Slack was slow or busy, test again later. When Hive hit a problem of its own, it can happen every time, for example when the Slack workspace has more channels or members than Hive reads in one check: waiting will not help, so contact support.
- Test run: not sent. A live run would be refused now: Shadowed: recipient not on record, not dispatched, for an email recipient that is not a valid address.
When an action’s result is unknown
A Hive update does not interrupt an action that is being sent: the step finishes first, and a deployed workflow’s run waiting for its next action continues on its own a few minutes later. A test run from the editor cannot continue after an update; it fails and you can test again. If an action’s result still cannot be known, for example because the machine sending it stopped mid-send, the step says prior dispatch outcome unknown — manual check required. Hive does not send it again and never bills it. Check the connected app to see whether it arrived.Cancel a run
Select Cancel run on a run that is still going. You can cancel only a run you started, and the button needs Can edit on the workflow. A run a teammate started shows no Cancel run button, even while it is still going. Actions already taken are not undone; check the audit log to reverse the reversible ones.When a run or deploy is refused
When Hive refuses to start a test run, or refuses a deploy, the reason stays in a row under the builder header in plain words, naming the step it is about, until you dismiss it. Show step opens that step. For example:
Before you deploy, the Deploy dialog lists every such problem at once under Before you
deploy; see Deploy a release.
Checks before a workflow runs
Hive checks a workflow before it runs, so a step that would fail is found before the run, not during it. The check looks at every step: whether its AI model is still available in this workspace, whether the tools and actions it uses can run here, whether a connector step has the fields its app needs, whether a saved agent it uses is still yours to run, whether your connected accounts are connected, and whether every value it uses comes from a step that runs before it.- In the builder. A problem only Hive can see — for example a step pinned to an AI model your workspace no longer allows — shows on the step’s card and in the setup checklist under the next step. On the card, a problem that stops the run has the card’s amber warning row, and one that still lets the run go ahead has a muted row; where Hive knows which setting to change, select the problem on the step’s card to open the step on that setting. The message names the step and the fix, such as “Send note” uses an AI model that isn’t available in this workspace. Choose another model for it. While your workspace has no AI model at all, the next step says so once, rather than on every AI step’s card. A draft with problems still saves. When the live version has a problem the draft has fixed, the checklist says The live version can’t run until you deploy.
- Problems that only stop going live. A problem that belongs to no single step does not stop you testing — for example a saved agent edited since the step chose its version, while the step still runs the version it chose. The builder then says Needs fixing before it can go live, followed by the problem, in the setup checklist and under Ready to go live in the next step. The Deploy dialog refuses the deploy until it is fixed.
- Run live now. If the live version has a step that cannot run, the run is not started, and a row under the builder header says This run wasn’t started, followed by the first problem.
- Schedules. A schedule cannot be turned on while the live version has a step that cannot run; Hive lists the problems instead. A schedule that is already on is checked again just before each scheduled run. If a step cannot run, that run is skipped rather than started: it is recorded in Past runs as a Failed run that starts Hive checked this workflow before its scheduled run and names each problem, and the schedule’s owner gets a Scheduled run skipped bell card — at most one a day for each workflow. The schedule stays on and runs at its next time once the steps are fixed.
409 with the code
workflow_connector_binding_required, and its details.nodeId is the id of the step to fix.
- An account refusal’s
detailsalso carryprovider(the app) andreason, which is one ofconnect,reconnectorchoose_accountand says what to do. - Deploy refusals, and (through the API and MCP) refusals to apply an assistant change or to let the
assistant build a workflow, use
connectwhen you have no account for the app,reconnectwhen the chosen account is gone or stale, andchoose_accountwhen you must pick one of your accounts. - Run refusals at start use
choose_accountwhen a step has no account chosen, even if you have no account for the app yet, orreconnectwhen a pinned account is disconnected or lacks permission. The pre-admission check used by public forms and schedules returns the latter asworkflow_connector_unavailablewithdetails.reasonset toreconnect.
On error: retries and continuing
Each step except the trigger has an On error choice:
Between attempts Hive waits 0.5 s, then 1 s, 2 s and so on, never more than 30 s, so with the
default 3 attempts every retry is over within about a second and a half. A step can allow up to 10
attempts when its policy is set in an exported workflow.
A Hive agent, Parallel agents or AI Condition step whose model is busy or rate-limited
works differently: when the provider says when to try again (its Retry-After), the step waits for
the model provider’s Retry-After when it fits: up to 2 minutes, or 30 seconds while a person, form
or MCP client is waiting on the run. Hive also pauses calls to a model for a while after its
provider turns Hive away, and a step waits out that pause the same way.
- Up to 2 minutes per wait and 2½ minutes in all for a run a schedule or a connected app event started, and for a run continuing after a Wait or Approval step.
- 30 seconds in all for every other run: one you start with Test run or Run live now, a form submission (in-app or public) and a run an MCP client starts, because someone is waiting for its answer.
- Each wait uses one of the step’s attempts. A wait that would go past the time left stops the step as below.
- When the model will not be back within that time, the step stops at once instead of spending its other attempts, and the run’s detail says when to expect it, for example OpenAI was busy; Hive expects it back at Nov 1 · 2:00 PM.
- A later step cannot read a step that may continue; Hive refuses to deploy if it does.
- Condition, If / else (rule), Approval and Output steps cannot continue: a failed condition decides nothing, and a failed approval was never approved.
- A refusal always stops the run, whatever the policy: an action above the role, a run that is not funded, a step your workspace cannot run, an app that needs connecting first, or an action a workspace safety rail stopped (see Why an action was not sent).
- An app that turns down one request is a failure, not a refusal, so a step set to continue carries on past it: Slack refusing a post (for example because the Hive app is not in the channel), or an app refusing a Read from an app step’s request. An app that needs a permission it was not granted, or whose sign-in no longer works (including Slack no longer accepting the connection), is a refusal and still stops the run: later steps would fail the same way until you reconnect it.
A connected app is unavailable
A step stopped as a connected app that is unavailable says what to fix:A step Hive cannot run
A connected-app step is refused, before any step of the run starts, when Hive has no reviewed action for what it calls. This happens when a step calls an app address directly, for example HubSpotPOST crm/v3/objects/tickets, and Hive does not run that call. It also happens when a step
uses an app action that is no longer offered. The message names the step and the call:
The step “Create HubSpot ticket” calls HubSpot POST crm/v3/objects/tickets, which Hive cannot run. Replace it with a HubSpot step from the catalog.Retrying does not help, and connecting another account does not either. Delete the step, add the app’s step from Add step in the builder, then deploy again. An app address starts without
/: Hive does not correct /crm/v3/objects/contacts, it refuses it.
Over the API and the Hive MCP server, the run, deploy and pre-admission
requests answer 422 with the code workflow_connector_step_invalid.
Hive also refuses to deploy such a step, and the deploy preview warns about it. The
workflow assistant does not propose one: it uses the app’s catalog
steps instead.
A Slack message checks its recipient against the people Hive knows: your
workspace members and, when Slack is enabled, the channels and members the run owner’s own
connected Slack account can see. If the run owner has connected more than one Slack account, Hive
does not guess which to read: only workspace members count as known, and a recipient a deployed
workflow fills in from its own form or connected app data is still allowed.
Slack being off, or connected twice, does not stop other actions. An email step can send to any valid email address (see Who an email step may send to). A draft, such as an Outlook or Gmail draft, sends
nothing, so it never needs that check.
The 500-step run budget
A run records at most 500 step executions. A step execution is one attempt of one step: a step that retries, or runs once per loop item, counts each time. This budget is per run. It is separate from the workflow’s credit limits, which count the credits its runs use. Hive counts the worst case when you deploy and refuses a workflow that could exceed it:- a retrying step counts once per possible attempt (3 by default);
- a Parallel agents step counts 1, plus each agent’s attempts;
- a loop counts 1, plus Max iterations times everything inside it, plus 2 more for each wait inside it on every iteration.
Credit limits
Each workflow’s runs draw AI credits: on Business, from the credits of the member the run is charged to (you for runs you start, the workflow’s owner for scheduled, app-event and public form runs; see Who pays); on an earlier plan, from the workspace’s pool. Three limits on the workflow’s Settings tab, under Limits, keep a workflow from using more than you expect. The credit limits count every run, test runs included; Runs per hour counts production runs:
Your workspace’s own limits still apply on top. On Business: the member’s own credits, their
top-up limit and the workspace’s monthly maximum on extra usage,
and the workflow runs the member and the workspace have
left this month. On an earlier plan: each member’s monthly limit, the
public budget for runs from a public form, and
the pool itself.
Confirming a large run
When you start a run that Hive estimates will cost more than £5 (500 credits on an earlier plan), Start this run? shows the estimate first, and the run starts only when you select Start run. On Business the estimate is in pounds, at the same fixed rate your credits are spent at; on an earlier plan it is in credits. Submitting a workflow’s form asks the same way. The run is then held to the estimate you confirmed for as long as it runs, including after a Wait or an Approval step and after Hive picks a run up again following a restart. Before you deploy, Deploy workflow also shows what each run is estimated to use at most, worked out the same way, so you know the cost before the workflow goes live. The estimate prices each AI step on the model it will run on: the model the step names or, for a step left on Workspace default, the workspace’s Default workflow model when an admin has set one, and Hive’s Efficient model otherwise. An Agent step that uses tools or reads connected apps is priced turn by turn, each turn at most 100,000 tokens in and 8,000 out, which is also what Hive sets aside for it while it runs.Runs past 50 agent steps
A run takes at most 50 agent steps without asking. Each Agent step, Agent library step, agent of a Parallel agents step, and each loop item that runs one of these, is one agent step.- A test run, Run live now or a form you submit that could take more than 50 asks first: Start this run? shows the most agent steps it could take and its estimate with every loop at its Max iterations. Once you select Start run, it runs to the end, held to that estimate.
- A scheduled run, one started by a connected app event, or one a visitor started from a public form, pauses before its 51st agent step and shows Awaiting input. Every Owner and Admin with a seat gets a card in their Inbox under Needs you: Approve lets the run carry on from where it paused, and Decline stops it as Rejected. If no Owner or Admin holds a seat, any of them can answer it under Billing notices in Settings → Billing, because it is a spending decision. If nobody decides within 72 hours, the run stops as Expired. A workspace with no Owner or Admin who could decide fails the run instead. While the workspace is read-only or paused, Approve is refused, because the run would act for the workspace; Decline still stops it.
Waiting for credits
When a scheduled run or a run started by a connected app event finds the workspace out of credits, it does not fail. It waits, and the workflow shows 1 run waiting for credits (or how many). The run starts by itself as soon as credits return, for example when an Owner or Admin buys a credit top-up.- A workflow with no AI step never waits. A run of a workflow with no Agent, Agent library, Parallel agents, Web search or Image step, and no condition that asks AI to choose, uses no credits, so it runs as usual while the workspace is out of credits.
- On Business, a run waits for its owner’s own credits. Each member has their own AI credits, so a run waits when the member it runs as (the schedule’s owner, or the workflow’s owner for a connected app event) has used theirs and extra usage cannot cover it. That member gets an Inbox card, Your workflow is waiting for AI credits, once for each waiting run. They can choose Request a higher limit in Settings → My usage, if the workspace allows usage requests; an Owner or Admin raises their own top-up limit instead, with Raise your limit in Billing.
- A schedule keeps one run waiting. While one of its runs waits, the schedule’s later times are skipped, each one recorded in the Audit log. When credits return, the workflow runs once to catch up, not once for every time it missed.
- A waiting run runs as the schedule would now: on the workflow’s live version and the schedule’s accounts. If the schedule was switched off or deleted meanwhile, by you or by a change of plan pausing schedules over its limit, the waiting run is dropped and does not run.
- A waiting run that starts runs once. If it then fails, it is a failed run like any other in run history, and it is not started again.
Timeouts
Hive has no overall run timeout. A step can run for as long as it keeps working: an Agent step that is still taking turns is never treated as stuck, however long it runs. A step that does nothing for two hours (a call that never answers) is treated as stuck. Hive then stops the stuck attempt and starts the step again: the stopped attempt takes no further turn and makes no further search or call, and if its call answers later, the answer is ignored, so nothing after it runs twice. If Hive restarts while a run is in progress (when Hive is updated, for example), the run carries on from the step it was on a few minutes later. The same happens if Hive cannot record for a few minutes that a run is still in progress (a database outage, for example): the step in progress stops at its next turn and starts again a few minutes later, so the turns it had already taken are taken again. When Hive starts a step again like this, because it was stuck, because Hive restarted or because Hive could not record that it was still in progress, the credits the stopped attempt used, its web searches included, come back to whoever paid for them (on Business, the member the run is charged to; on earlier plans, the workspace’s pool), so the step is paid for once. That includes what the stopped attempt used after Hive started the step again, if it took a few more moments to stop. If the run then fails on Hive’s side, its refund and these credits never give back the same credits twice, and these credits never count towards the 5% monthly limit on refunds, even if the run’s refund came first. If the database is still unavailable at that moment, Hive keeps trying to give the credits back for about half an hour after that. If it stays unavailable for longer, some of these credits may not come back on their own: contact support if a run’s credits look wrong. Two steps never pay for the same work twice, so there is nothing to give back:- A Generate image step whose stopped attempt had already saved its picture to the Library uses that picture instead of drawing another. If the picture was not saved, the step draws it again and the stopped attempt’s credits come back.
- An Agent library step never runs its agent twice. It uses the answer the agent’s run already gave, or fails if a restart cut that run short.
When runs do not start
An unattended run (a schedule, a public form or a connected app event) is refused when:- the workflow is paused or not deployed;
- the owner no longer has the Operator role;
- the owner holds no Basic, Pro or Max seat: their seat was moved to someone else or suspended over the plan’s limit. Each run is skipped until they have a seat again, and Owners and Admins get one notice — see When someone loses their seat;
- the kill switch is on;
- the workspace is read-only, because its trial ended without a plan, its plan was cancelled or a payment has been overdue for 14 days, or a payment is overdue (schedules and connected app events stop while it is) — see Trial and payments;
- the workflow reached its Monthly credit limit or its Runs per hour;
- the plan does not include the trigger (connected app events need a paid Business or Team plan; they are off during the 7-day trial), or, on Team, the workspace used its plan’s connected-app-event runs for the billing period;
- on Business, the owner’s own workflow runs and the workspace’s execution add-on are both used for the month — see Workflow runs;
- a public form reached its 500 runs for the day or 5,000 for the month;
- a public form run found its owner, or on an earlier plan the workspace, out of credits;
- a connected app event was too large for a run to record — see Refused unattended runs.
Related
Test and deploy
Test runs, releases and notifications.
Triggers and steps
Every step and its settings.
Limits reference
Every workflow limit in one place.
Audit log
Receipts and undo for actions taken.