Skip to main content
Agent Studio is where you build specialised agents you configure for a job — grounded in your Business Brain, held by your rules. Ask Hive is a general assistant. An agent is a saved, versioned specialist, such as a collections chaser that writes in your tone, a proposal writer that knows your templates, or a triage assistant for one shared inbox. In the sidebar, select Agent Studio (/agents), just above Workflows. The page heading is Agents. The command palette (⌘ K) also offers it.

What an agent is

An agent combines five things, all saved together as a revision: An agent reads its own attached files for everyone who runs it, and otherwise only what the member it runs as can see. It takes only the actions selected for it. Hive treats the persona and task text as untrusted, so nothing in an agent’s instructions can widen its permissions or add an action.

Attach files as knowledge

In the agent editor’s step 2, Knowledge & tools, the Files row lists the Library files the agent reads (see How the editor is laid out). To add one:
  • Select + Add files and pick files that are already in your Library, or open Upload files in the same picker to upload new ones.
  • Open Google Drive, OneDrive or SharePoint in the same picker to add a file from your own connected account: search your Drive and OneDrive files by name, starting from the ones that changed most recently, or browse a SharePoint site, its document libraries and their folders. See Files from Google Drive, OneDrive and SharePoint.
  • Drag files from your computer onto the Files row or onto the Build composer. A dashed Drop files to add to this agent outline shows where they will land.
Uploads go to your Library as private files and take the usual checks: up to 10 files at a time, each a supported type and no larger than 25 MB. You need the Operator role or higher to upload. Each upload appears in the Files row straight away. It reads Processing… while Hive reads its content, then Attached · Save to keep once the agent can use it. Save the agent (Save draft, or Publish changes for a live agent) to keep the attachment as part of a new revision; until then, Discard changes removes it, and the file stays in your Library. If an upload fails, its chip shows the reason with Retry and Remove. The same happens if a file is still processing after 10 minutes, or if Hive can’t check its status 10 times in a row. Select a file’s name in the Files row to open it in the same preview Library uses: a Word or PowerPoint file opens on its pages, with Text one click away. The preview opens under your own Library access, so a file you can’t open in Library shows no open action here, only × to remove it. Opening a file never changes the agent.

Files from Google Drive, OneDrive and SharePoint

A file you add from Google Drive, OneDrive or SharePoint is copied into your Library, then attached like an upload (Attached · Save to keep). The agent reads the copy, never the live file, so a run never waits on the original or fails because of it. You need the Operator role or higher, and the account must be one you connected yourself in Integrations; if none is connected, the picker says so. In the Files row a copy says where it came from and when Hive last checked it, for example Copied from Google Drive · checked 2 h ago. Only the person who added it can refresh it:
  • Select Refresh on the file to check the original now.
  • When you open the agent in Agent Studio, Hive checks your copies whose last check is more than an hour old.
  • When you deploy the agent, Hive checks your copies again.
Nothing is checked in the background, and nobody else opening the agent starts a check. When the original has changed, Hive copies it again and keeps the previous content in the file’s version history. The agent uses the new content straight away, deployed versions included, as it would after an edit to a Library file. If a check can’t reach the original, the copy keeps working and the Files row says why, such as Couldn’t refresh: reconnect Google Drive. See Files copied from Google Drive, OneDrive and SharePoint.

Who can read an agent’s files

The files and folders attached to an agent are part of the agent. When you share it, everyone who can run it gets those files when they run it on its run page, in Ask and in Continue in chat, even files only you can open in the Library:
  • In its runs. The agent reads its attached files on its run page, in an Ask chat where someone picks it, and in Continue in chat. Hive reads them with the Library access of whoever saved the agent’s latest version.
  • Pictures. When the agent’s model reads images (a vision model), an attached JPEG, PNG, GIF or WebP picture reaches it as the image itself on its run page, in Ask and in the editor’s Test. A model that does not read images gets Hive’s description of the picture instead (see How Hive reads the files you choose).
  • In its chats. The person running the agent can open a file the agent used from that chat, read-only: select the file’s name under Sources. A PDF or an image opens in a new tab; any other file downloads.
  • Nowhere else. The files are not added to the runner’s Library, its search or Drive, and they can’t open them anywhere outside a chat with the agent. Opening the original in Google Drive, OneDrive or SharePoint works only if that service lets them.
  • Removing a file from the agent and saving it stops it straight away: new runs no longer read it, it no longer opens from the agent’s chats, and earlier answers in its chats that used it are hidden. A run’s own result on its run page stays visible to the person who ran it.
  • If the person who saved the agent’s latest version leaves the workspace, or can no longer open one of its files themselves, the agent’s files are read only for people who can open them in their own Library, until someone who can open them all saves the agent again.
  • Someone your workspace keeps out of the Library can’t use an agent’s files, as before. An AI client connected to Hive’s MCP server can’t run agents or open their chats, so it never reads an agent’s files this way.
Files a person chooses for one run (a run source, or a file in a form field) are still read with that person’s own access, and so are the files of an agent that another agent hands a task to, and of a run started with Run now, on a schedule or from a Signals hand-off. A workflow step reads an agent’s files as the member the workflow runs as, and a public link reads them as the person who published it.

Where agents run

Every deployed agent is available in Ask, on its own run page and in workflows. You don’t choose where it appears. A public link is separate: you turn it on from Public link in the editor.

Ask Hive

Pick the agent in the Ask composer’s agent selector instead of the default Hive.

Its own run page

Fill in the agent’s form, run it, read the answer and see its past runs.

Schedules

Run the agent on a recurring schedule with a fixed prompt and timezone.

Workflows

Add a deployed agent to a workflow from the builder’s Agent library.

Public links

Share an agent outside the workspace, if your workspace has this turned on.

What an agent can use where it runs

Your workspace’s mandatory skills join the agent’s instructions in each of these places except a public link.

Use an agent in workflows

Deploy the agent. Its deployed revisions then appear in the workflow builder’s Agent library, under Agents in the Add step menu and in the Add a step rail. A workflow step pins the revision it was added with; when you deploy a newer one, the step offers Update to revision N. A workflow step reads only Library files and folders or Google Docs links as knowledge and run sources, so the Agent library greys out an agent with other knowledge, skills, sub-agents or a connector operation that writes, or with other sources, and says what to change. Web access, The person’s own data (Hive data included) and read-only connected apps work in a step: see Web, Hive data and connected apps. Its selected actions run in the step. An agent built before workflow steps existed must be rebuilt from a reusable template. See Agent library steps.

Run an agent directly

Every deployed agent has a run page of its own at /agents/<agent id>/run. Open it with Run in the agent card’s ⋯ menu, or with Run this agent and see its runs in the editor’s Test pane. You can use it if you’re an Operator or above with run access to the agent: you don’t need to be able to edit the agent, and the page never shows its draft versions. The page shows the agent’s name and description, then its form:
  • Fields: the inputs the agent asks for.
  • Sources: the Library files and folders, or other sources, the agent reads for this run. When a source takes Library files, you can add a file that isn’t in the Library yet with Upload to Library in its picker; Hive selects it for you once it’s uploaded. You need the Operator role or above. Each file can be up to 25 MB and of a type the Library accepts. It goes to the root of your Library and follows Library sharing, like any other upload.
  • Accounts: which of your connected accounts the agent uses, when it works with connected apps.
  • Run instructions (optional): a note for this run.
  • Conversation starters: the agent’s starters, listed under Run instructions (optional). Select one to put it in the run instructions; it doesn’t start the run.
An older agent saved before run setups existed shows a simpler form instead: its fields, Run instructions (optional) and Run now, without uploads or Continue in chat. Select Run. The answer appears below the form, answer first, with its sources and any files it made folded underneath. When your run instructions ask for a chart, such as “chart revenue by region”, the agent draws it from a spreadsheet attached to it or one it reads, or from numbers you typed, and the chart is shown and saved with the answer, its source named underneath, as in Visualize data. A run started from Agent Studio — Run now, a schedule, an app event, a signal or another agent — keeps the charts it drew too, up to four, shown under its answer when you open the run; a chart drawn from sample numbers is labelled as sample data, and its answer says so. Charts are shown and withheld exactly as the run’s answer is. If the agent allows follow-ups, Continue in chat opens the run as an Ask chat so you can carry on the conversation, building on up to 50 earlier answers; after that, Hive asks you to start a new session. If it does not, each run gives one answer: run it again to ask something new. A run starts as soon as you select Run. Hive reads its sources once, with your current access, when the agent works on the answer: if you can no longer open a source, the run fails and says so. A workflow’s saved-agent step works the same way. If the run held any actions because of where they send, its answer says how many are waiting for your approval under Review decision on your Signals page.

Files a run writes to your Library

An agent with a Word or Excel file template can write the document or workbook it promises, such as an engagement letter or a proposal, as a new file in your Library when you run it yourself on its run page:
  • Your file. The file is yours, written as you: at the top of your Library, or in one of the folders attached to the agent that is private to you and that you can add files to. A folder shared with anyone else, and a Knowledge Base folder, are never used, so what the agent read for you never lands where the agent’s author or a colleague could open it.
  • On the agent’s template. A Word file is written on the agent’s Word template and a workbook on its Excel template, when you can open that template. A private template is used only for a file that stays private to you; otherwise the file takes your workspace’s own design. On one of Hive’s starter templates, the header and footer name your firm in place of “[Company name]” and “[Firm name]”: the name on your firm record, else your workspace’s name. If Hive knows neither, the placeholders stay in the file for you to fill in.
  • New files only. A run never changes, replaces, moves or deletes a file that is already in your Library. It writes at most three files, each a draft for you to check.
  • In the result. Each file is listed under Artifacts in the run’s result, with Open, which opens it in your Library, and the answer links it too.
  • Recorded. Each file a run writes is recorded in the audit log as “Agent run created a Library file”, with the agent, the run and the file.
A workflow step’s run (including a scheduled or app-event run), an author’s Test and a public link never write files: use a workflow’s Create PDF step to file a PDF from a workflow. You need the Operator role or above and access to the Library, as for any file you add. If the agent’s AI model fails partway through a run, the run says which model failed, what its AI provider reported (for example that it rejected the request, with the HTTP status and the provider’s error code), and what to do next: wait and try again when the provider is busy, change the request when a safety filter blocked it, or try again and then choose another model in the agent’s settings when the provider keeps rejecting it. The same message appears in Agent Studio, in Ask and on a workflow’s saved-agent step.

Run history

Below the form, Run history lists the agent’s runs, newest first:
  • You see your own runs. Workspace admins see everyone’s runs and can filter by person.
  • Each run says who started it: You for your own runs, the member’s name on someone else’s, or Workspace member for someone no longer in the workspace. The Recent runs list on the agent’s page in Agent Studio shows the same, on each row and in Started by when you open a run.
  • Filter by status. Load more shows older runs.
  • It includes runs started on this page, in Ask, by a schedule, by a workflow, and older runs started with Run now.
  • Each run has a link of its own, so a notification about a run opens it here.
  • A run a limit stopped shows as blocked, with the reason: the workspace is out of credits, a member’s or the agent’s monthly limit is reached, the workspace is read-only or has an overdue payment, or the model policy does not allow the agent’s model. See Credits and usage.
  • What was asked and the start of the answer show only on your own runs. On someone else’s run, an admin sees who ran it, when, how it started and its status, but not what was asked, the answer, its charts or its steps. This applies to the run page, run history and the recent runs on the agent’s page. An agent’s answer in an Ask chat is part of that chat, so anyone who can open the chat sees it, including an admin who can read members’ chats.
  • A run whose Ask chat was deleted is read only: you can open it and read its answer, but not continue it.

Browse and organise agents

The list shows agents you own and agents shared with you. Admins can also see agents across the workspace.
  • Sort by Most popular, Recently updated or A – Z.
  • Each card shows how many times the agent ran in the last 30 days and when it last ran. Every way it runs counts: Run now, schedules and triggers, its own page, Ask and workflow steps. You count the runs you started; an admin counts everyone’s. Most popular sorts by the same runs.
  • Each card names the agent’s owner: You, or the member who made it. An agent Hive set up in your workspace for you (see Suggested for your workspace) shows Hive as its owner, never a person’s name.
  • Filter by status: Live, Draft, Paused or Archived.
  • Search agents by name and description. Pin the ones you use most so they appear under Pinned agents.
  • Import agent brings in an exported agent as a new draft. Browse templates starts you from a ready-made agent made by Hive (see Templates by Hive). + Create agent opens the builder, where you can type or dictate what the agent should do (see Start a new agent).
  • When a tab has no agents yet, it says so and offers Create agent if your role can create one. A search or filter that finds nothing asks you to clear it instead.

Quick actions on an agent card

Each agent card has a ⋯ menu, so you can act on an agent without opening it:

Agent statuses

Who can do what

Agent Studio follows your workspace roles (see Roles and permissions):
  • Create an agent: Department lead, Admin or Owner. Anyone else sees “Only a lead, admin, or owner can create agents.” On Starter, a workspace can have 3 agents that are not archived, counting every member’s; Team and above have no fixed limit. Creating or copying one more at the limit is refused, and a workspace above its plan’s limit (after a downgrade, for example) cannot deploy an agent until it archives some.
  • Run an agent: Operator or above, plus run access to that agent.
  • Delete an agent: Department lead or above.
  • Manage workspace agent settings: Admin or above.
  • Share: an agent is private to its owner until they share it with named collaborators. Sharing an agent never shares the owner’s connected accounts. It does let everyone who can run it use the agent’s attached files when they run it on its run page, in Ask and in Continue in chat, and open them from its chats, without adding them to their Library: see Who can read an agent’s files. Its name and icon still appear on the answers it gave in an Ask chat, for everyone the chat is shared with; see Use an agent in Ask.
Admins can turn Agent Studio off for the workspace, or limit it to selected members and groups, in workspace settings.

Safety built in

An agent’s actions follow these rules:
  • Selected actions follow the account’s level. An agent’s selected action runs from the member’s own account, and that account’s level — with any workspace ceiling applied — decides it: on Act it runs straight away; on Draft it waits on an approval card; on Read it takes no action and the answer says so. That holds however the run started: by a member, a schedule, an app event or a signal. See What agents and workflows do by themselves.
  • Where an action sends. An action whose recipient, channel, phone number or link did not come from the agent’s instructions or from what the member typed in Hive for that run, such as an address the agent found in an email it read, waits for that member’s approval as a card instead of running: in the chat, or under Review decision on their Signals page. A schedule’s saved input does not count as anyone’s own words. Write the addresses, channels and links the agent may use into its instructions yourself: Describe it and the Build chat never add one you have not saved or typed in your description, and say so instead (see Build an agent). See Where an action sends.
  • Actions on or off. Set What it may do to Only reads and answers under Advanced in step 3, Inputs & output, and the agent takes none.
  • Selected actions only. An agent takes only the actions selected for its attached apps. Hive refuses any other action before it reaches an account. An agent set up before actions were chosen per app takes none until you select them.
  • MCP tools for everyone who runs it. A selected MCP tool runs on the runner’s own connection to the same server. A runner without one is told to connect the server (or, when Hive couldn’t load its tools, that it couldn’t); a runner with two connections to it is told to disconnect one. A tool the server has changed since it was ticked is refused with a note asking the author to tick it again in Agent Studio, and the answer notes any tools ticked for their author’s own connection only. Only tools from servers the editor has connected can be newly ticked.
  • Budgets. Hard limits on run time, model calls, action steps, external writes and tokens. A run takes at most as many turns as its model-call budget, and never more than 16. New agents start with Hive’s defaults (8 model calls), an agent made from a template that chains several tools starts with more, selecting an agent’s first action allows one external write per run, and agents built earlier keep the budgets they were saved with. Past the write budget, Hive refuses further actions. An action that waits for approval because of where it sends uses none of the write budget, but counts as one of the run’s action steps. Your plan’s token ceiling for one run applies on top, in Ask too.
  • Workspace limits. The hourly action limit and the spend caps still apply to actions. An action they stop is refused, and the agent says why.
  • Credits. An agent’s model work, whether it runs here, on a schedule or when someone asks it in Ask, uses AI credits: on Business, the credits of the member who runs it (a scheduled run, the schedule’s owner; a public link, its publisher), with no monthly limit per agent beyond that member’s own credits and top-up limit; on an earlier plan, the workspace’s shared pool, up to a monthly limit per agent of 2,000 credits on Starter and Team and 10,000 on Business and Enterprise. If you raised an agent’s token budget above the default, a run Hive estimates at more than £5 (500 credits on an earlier plan) asks you to confirm before it starts; an agent on the default budget never asks. The Deploy dialog shows what each run is estimated to cost at most, before the agent goes live: in pounds on Business, in credits on an earlier plan. A scheduled or event-started run whose credits have run out waits up to 72 hours for credits, and the agent’s card in Agent Studio shows 1 run waiting for credits (or how many), or Run when credits return? once a run has waited 72 hours (see Waiting for credits).
  • Repeated failures. An action that fails three times in an hour on the same account pauses on that account only, and the agent says until when. Reconnecting the account lifts the pause. See When an action keeps failing.
  • Outside content. An agent started by an app event (a new email, a new deal) reads what someone outside your workspace wrote and still acts. An action that would send to an address, channel or link taken from that content waits for approval, but an action on an existing record, such as replying on a helpdesk ticket, sending or voiding an invoice or deleting a contact, is not held this way. Select only actions you are happy for it to take on such content, and keep its write budget small.
  • Kill switch. When an admin turns on the kill switch, agents stop acting along with everything else. A run or Test already in progress stops before it takes an action, reaches outside Hive (a web search or a connected app) or saves its answer. Until then it may make further model calls, which are charged as usual, but nothing it produces is saved or delivered. Pausing, archiving or un-deploying the agent, or removing the person who started the run (or moving them below Operator), stops a run in progress at the same points.
  • Receipts. Completed actions are recorded in the audit log.
You can still see a saved run’s status and stop an unfinished run if its Library source becomes unavailable. A finished run keeps its answer: you can open it later even if you have since lost access to a source it read. Losing access applies from your next run or question, which no longer reads that source or cites it.
Reusable member agents are off by default. You’ll only see them if Hive has turned them on for your workspace. Public agent links are available to every workspace: see Public agent links.

Limits at a glance

See the limits reference for every limit.

Build an agent

Describe, configure, test and deploy your first agent.

Publish and evaluate

Analytics, feedback, schedules and public links.