Skip to main content
This guide walks you through building an agent in Agent Studio, from a one-line description to a deployed revision your team can use. You need the Department lead, Admin or Owner role to create agents.

How the editor is laid out

The Setup tab has two sides. On the left, Build is a chat with the builder and Test tries the agent. On the right, the agent’s setup is split into four steps:
  1. Basics & instructions: the name, description, icon, labels, model, instructions and conversation starters.
  2. Knowledge & tools: files, folders, integrations and their actions, skills, sub-agents and file templates.
  3. Inputs & output: how people start it, form fields, files people add each run, what it returns, what happens after the result, and Advanced.
  4. Review & deploy: a summary of every step, what is still missing, and the deploy action.
A small step bar above the steps shows Back, the four steps, Step 2 of 4 and Next. On the last step, Next becomes the deploy action. You can open the steps in any order, and nothing is locked. Each step has its own address (?step=knowledge), so you can link to it, and the browser’s back button moves back one step. Moving between steps never saves and never loses an edit, even one you have half finished. An agent you open from the list starts on Review & deploy. A new agent you described or made from a template starts on Basics & instructions. On a phone, the editor fills the screen. Setup, Build and Test show one at a time, a thin bar under the agent’s name shows which step you are on, and Back and Next sit at the bottom. Save draft joins them once you have changes. Analytics, Feedback, Past chats and the other agent actions are under the ⋯ button.
1

Start a new agent

Open Agent Studio in the sidebar and click + Create agent. You have three ways to start:
  • Describe it. Write what it should do in the box (“Describe what it should do…”), choose the new agent’s Agent model from the model picker (the same one as in Ask; it starts on your workspace’s Default agent model when an admin has set one), and Hive drafts the agent for you. For example: Chase overdue invoices in our tone and log every touch, Triage Gmail and draft replies, or Turn notes into client-ready proposals. To speak instead of type, select the microphone next to the send button. What you say is added after any text already in the box, a phrase at a time as you pause. While you record, the microphone becomes a recording bar with a level meter and the time so far: select the tick (Insert text) to stop and keep what you said, or the cross (Discard recording) to put the box back as it was. Dictation stops by itself after 10 minutes and keeps what it heard. On Business, dictation uses your own AI credits, at what the transcription costs Hive; on an earlier plan, each Pro or Max seat adds 5 hours a month to the workspace, and after that it uses 1 credit a minute.
  • Choose a template from the template cards under the description box, or from the full list under Templates by Hive.
  • Create draft manually to start from an empty agent.
  • Ask for it in chat. In Ask, say “make an agent that summarises my unread Outlook email each morning”. Hive drafts it on the chat’s model and shows a card; Create saves it here as a draft. See Build from chat.
A generated agent opens as a reviewable draft on step 1, so nothing is live until you deploy it.Every new agent starts on a model your workspace can run: the Default agent model when an admin has set one, otherwise your workspace’s default model. You can change it under Model in Basics & instructions.Create draft manually opens the new agent as a conversation instead, like briefing a new co-worker: “Hi, I’m your new co-worker”. Tell it, or say it with the microphone, what you want taken off your plate, or tap one of the three ideas. Each answer is a request to the builder, which proposes the changes for you to build in. Beside the conversation, What I know so far ticks off Named, Has instructions, Has a way to start and Has knowledge or tools, with Deploy under them. Set it up step by step opens the four steps.
2

Write the instructions

In step 1, Basics & instructions, give the agent a name, description and avatar, then write its Instructions: the job, the tone, what “done” looks like, and what it must never do. The base prompt can be up to 16,000 characters.You can build and change the agent by chatting with the builder in the Build pane on the left. Before your first request, the pane shows the Hive disc, Edit with a sentence and three things you could say, such as Add two conversation starters: tap one to send it. You can also speak a request with the microphone in the box. Describe what you want, for example “add Gmail, switch to a cheaper model and add two starters”. While the builder works, the pane shows your request as sent with a pulsing Thinking about your change… row under it, the suggestions are hidden, and the send button becomes Stop builder. The builder works like this:
  1. It asks first when something is unclear, such as which mailbox or which files, with up to three short questions.
  2. It proposes a plan: the parts of the agent it will change, such as Model, Apps and actions or Conversation starters. Review proposed changes shows the exact values.
  3. Start building applies the plan to your draft in the editor beside the chat. Each changed field is highlighted and the chat lists What changed. Show in step N opens the setup step that holds the change, and Undo puts the draft back as it was before, as long as nothing has changed since. In the new-agent conversation, Try it first opens the Test pane and Review the setup opens step 1. Ask for edits sends you back to the chat to adjust the plan instead.
  4. You can still edit any field by hand. Nothing is saved until you save the agent (see Save and deploy).
The builder can change the agent’s name, description, icon, instructions, model, capabilities (web search, image generation, creativity and member data), Library files and folders, apps and their actions, skills, sub-agents, form inputs, conversation starters and labels. It only offers what you could pick yourself: models you may use, apps your role allows, your own Library, skills you can attach and deployed agents you can open. When you ask for a cheaper model, it picks one that uses fewer credits: Efficient models use the fewest, then Balanced, then Frontier. A sub-agent with required inputs must be added by hand, so you can map them. The builder never changes sharing, public links, schedules or run budgets. Read each plan before you build it.Two changes are always yours to make:
  • A new action. Selecting an action is a real permission: the agent carries out the actions selected on its apps, through the account of whoever runs it. So the builder never selects one. It can attach an app, keep its actions or remove one, and when your request needs a new action it tells you which one to switch on yourself under that app’s Actions (see Connect its tools).
  • A new place to send. The builder never adds a recipient, channel, phone number or link the agent does not already have, even one you typed in the Build chat, since an AI client connected to Hive can write in that chat too. It leaves the value out and tells you to type it into the instructions yourself. Your own connected mailbox’s address always counts, so “email me a summary” works. Describe it follows the same rule for a brand-new agent: a draft may name only an address, number, link, channel or repository you typed in your description, plus your own mailbox. Otherwise Hive creates nothing and says, for example, “This agent’s draft names finance@acme.example, which you did not type in your description. Describe the agent without it, then add it to the instructions yourself.”
Describe it and the Build chat run on Hive’s Efficient model (your workspace’s Default background model, or else Hive’s default Efficient model), whatever model the agent itself uses; there is no builder model to choose. They run through that runtime, with the same allow-list and billing as Ask: builder turns use credits from your workspace’s shared pool, at Efficient rates, and count toward your monthly credit limit, where one applies. Every change is checked on the server against the agent definition and the action catalog before you see it. A plan made against an older saved revision is refused instead of applied.
3

Choose how people start it

In step 3, Inputs & output, pick how people start it:Add up to 12 Conversation starters, which are one-click example prompts shown to people using the agent: under the box when they choose the agent in Ask, under the run instructions on its run page, and under the message box on its public link before a visitor’s first message.
4

Ground it in your knowledge

Attach Library files and folders the agent should read, such as policies, templates, price lists or past proposals. Attached files reach the agent as material to work on, at the start of the conversation, not as its instructions, so text inside a file cannot change what the agent does. If the agent must follow a file, such as a refund policy or a price list it must quote from, say so in the instructions (“follow the refund policy in Refunds.pdf”). Attached skills still count as instructions. Everyone who can run the agent gets these files when they run it on its run page, in Ask and in Continue in chat, and can open them from its chats, even if only you can open them in the Library; they are never added to anyone else’s Library. See Who can read an agent’s files. An answer the agent gave in a shared Ask chat is the exception: everyone the chat is shared with sees it in full, though not the files behind it (see Feedback, sharing and history). In a Library with more than 100 files, the file picker opens on the 100 most recently updated. Search by file name to find any other file, or click Show more files. To add a file from your own Google Drive, OneDrive or SharePoint, open Google Drive, OneDrive or SharePoint in the same picker: Hive copies it into your Library and attaches the copy (see Files from Google Drive, OneDrive and SharePoint). See Library.Under More → Browse templates, choose Use … for this agent’s files on a Word, PowerPoint or Excel file template (up to six). When a run of the agent writes a document or deck, it is generated on the template of that kind, keeping its layouts or styles; a workbook follows the template’s sheet names and header rows. The template is referenced, not copied: whoever runs the agent must be able to read it, or the file is written without it. Select Stop using … to remove it, then save the agent. With a Word or Excel template, a run someone starts on the agent’s run page saves the document or workbook as a new file in their own Library (see Files a run writes to your Library).The same picker can create a private reference copy of a file template, or Generate a reference from a Word or PowerPoint template, to read as knowledge. Describe the new content, then select Generate and attach to stage the generated file as knowledge. The content is written on your selected Ask model rather than the deployment’s default model. Review the fidelity warnings in the picker, use Open generated file in Library to inspect its content, then choose Done reviewing and save the agent to keep the attachment. This generates the file once while editing the agent; agent runs read the resulting file subject to the caller’s access. If the new file isn’t ready yet or you lose authoring access, it remains in Library without being attached.Under Skills, choose + Add skill to attach up to 10 skills: your own, the workspace’s shared and mandatory skills, and enabled skill packs. Every run of the agent reads them in full, as the member who runs it, so a personal skill applies only to your own runs; share it with the workspace for everyone else. Mandatory skills apply without being attached, as the member the run reads as: in Ask, the agent’s run setup, Studio Test and workflow steps. A public link’s runs do not use them.For a different file or folder on each member run, declare a Library Run source instead. The runner chooses it from sources they can currently access. In Ask, the agent’s form starts with the chat’s Library folder or file (or else the chat’s project folder) chosen when it is offered ready for a compatible run source, and nothing else; the runner can clear or change it, and Use chat folder or Use chat file chooses the chat’s one again. Otherwise search for an accessible source. Only the source shown in the form is used: the agent never reads the rest of the chat’s Library context. Information a workflow passes in declares a separate input that a workflow hands the agent; it is never taken from the chat. In the editor, run sources are listed under Asks for a file or doc each time.
5

Connect its tools

Under Integrations, click + Add integration and choose one of your connected apps, for example Gmail, Outlook, HubSpot or Slack, or one of your MCP servers. Select Attach to let the agent read it. Reads need no other setting. An MCP server is listed even when it has only reads and no action to select, so an agent can be given its reads alone.An agent with many apps attached doesn’t send the model all their tools on every step: as in Ask, it starts with its everyday tools and searches for the rest when a run needs them, only ever among the tools the agent was given. See How Ask picks its tools.The same screen lists the actions Hive can take in that app, each with an on/off switch, for example Send email in Gmail or Outlook, Create draft in Gmail, Create deal, Update contact and Create note in HubSpot, Post message in Slack, or an MCP server’s write tools. Each action is listed once. Switch on the ones the agent may take: each action that is on shows Runs automatically when someone runs this agent. There is no “ask me first” setting, because starting the agent is the permission, apart from the check on where an action sends described below. Switching an action on attaches the app for you, turns the agent’s actions on and allows one external write per run. An MCP server’s write tool counts against the same per-run write budget as an app’s action. Clearing the last action turns actions off again. An attached app you haven’t connected yourself shows Connect on its chip, which opens Integrations in a new tab. The list shows only actions Hive can carry out and that your role may propose. The agent can’t take any action you didn’t select. If it tries, Hive refuses before anything reaches an account, and says so.Every run uses the account of the member who runs it (for a schedule, its owner), never yours, and that account’s level decides each selected action, with any ceiling an admin has set for the app applied, as in Ask and Signals: on Act it runs automatically, on Draft it waits for the member’s approval, and on Read it is not carried out. A few actions still wait for a click even on Act, such as a LinkedIn post. The other check is where an action sends: one 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 waits for their approval as a card instead. A selected Send email can go to any valid address, but one the agent found in something it read, such as an incoming email, waits for approval first. See What agents and workflows do by themselves.
Write where the agent sends into its instructions, for example “post the summary to #finance-alerts” or “copy accounts@yourfirm.com”. Write the address itself, not only a person’s name. Name a Slack channel as #name, a Telegram chat as @username and a repository by its GitHub link: a channel, chat or repository named only by an ID waits for approval each time, and so does a Microsoft Teams channel, however you name it.
An MCP server’s write tool is selected for the server and the tool as it was when you ticked it, not for your own connection, so it runs for everyone who runs the agent, on their own connection to that server. You can tick only the tools of servers you have connected yourself; a tool already ticked on the agent stays when someone else edits it. Copying an agent, or restoring a version that brings back a tool since unticked, needs the same: connect the server first. When it runs:
  • Someone who hasn’t connected the server is told to connect it in Integrations. If they have, but Hive couldn’t load its tools for that run, they’re told that instead.
  • Someone with two connections to the same server is told to disconnect one, because Hive can’t tell which account you meant.
  • If the server changes what the tool does, the agent stops using it and says so until you tick it again.
  • A tool ticked before this worked for everyone runs only on your own connection and shows Selected for its author’s own connection only: tick it again to let it run for everyone.
  • A ticked tool whose name plainly reads, such as get_current_user or list_accounts, runs as a read while the agent works, so the agent uses what it returns in its answer, as long as the agent may take actions. Any other ticked tool is carried out as one of the agent’s actions.
If Hive couldn’t load your MCP servers’ tools when you open the list, it says so and offers Load MCP tools again, rather than showing none.An agent saved before you could choose actions per app takes no actions until you select them. The editor says so under Integrations.A saved agent that declares a Google Docs Run source needs an available Docs reader and the runner’s selected account. If that reader is unavailable, Hive refuses deployment. Library files and folders use the runner’s existing permissions and do not require a connected Docs account.
6

Set its behaviour and limits

In step 3, Inputs & output, open Advanced, then under Agent behaviour choose What it may do:
  • Only reads and answers: it looks things up and replies, and changes nothing.
  • Can do things in your apps: it also carries out the actions you switched on under Integrations.
With actions on, each selected action follows the level of the runner’s account: on Act it is carried out straight away, up to the agent’s external-write budget, for anyone allowed to run the agent — including runs a schedule, a connected app event or a signal starts. On Draft it waits for the runner’s approval, and on Read it is not carried out. On Act, an action whose recipient, channel, number or link did not come from the agent’s instructions or the runner’s own words waits for the runner’s approval instead, and uses none of the write budget; a schedule’s saved input does not count as the runner’s words. The workspace’s kill switch, hourly action limit and spend caps can still stop one; Hive then refuses it and says why. An action that fails three times in an hour on the same account pauses on that account (see When an action keeps failing).Also under Agent behaviour, set these with switches and buttons. The (i) next to a setting explains it in more detail.
  • Search the web: public web searches only.
  • Make images: the agent may draw a picture when a request needs one, in a saved chat, in its runs and when another agent hands it a task. Each image is charged. The workspace image setting still applies. Off by default; without it, a member can still ask for one with the image option in Ask.
  • The person’s own data: switches for which of the member’s own Documents, Business Brain, Signals, Hive data and Connected systems the agent may read, as that member and within their access. Test and a sub-agent’s run read them the same way. Connected systems includes the member’s own matters: the agent can read one with its requirements and timeline. To mark a requirement done or not needed, use the matter’s Requirements or ask Hive in Ask. An agent with apps attached under Integrations sees only the matters, and the track record of past actions, that come from those apps: an agent with Outlook attached finds a matter built from Outlook mail, but not one that also draws on HubSpot. Leave them all off for an agent that answers from its own knowledge only. Public links never get member data.
  • Creativity: Precise, Balanced or Creative. Use recommended picks a steadier setting when the agent has knowledge or skills attached, and then shows Recommended. An agent you haven’t set this for uses Hive’s default. Any other value set before these three choices existed shows as Custom and is kept until you pick one.
  • When credits or limits run out: Stop, or Switch to a faster model. Switch to a faster model moves the run to Claude Haiku when the agent’s own model can’t take a step: it is rate-limited, overloaded or unavailable, before the step is sent or when the provider turns the step away without doing any of it. The rest of that run then stays on Claude Haiku, and each step is charged at Claude Haiku’s rates. Whichever you choose, a run whose models can’t take a step then moves once to another model your workspace allows, the way Ask does (see Models for agents). A step the provider started on and then failed is not sent again on another model. Running out of credits always blocks the agent, whichever you choose; it never switches the model.
Run budgets aren’t shown in the editor. New agents start with Hive’s defaults, and an agent built earlier keeps the budgets it was saved with.
7

Test it

Try the agent before anyone else uses it:
  • Draft preview runs the unsaved draft. It never writes to your tools; it reads only what the draft’s own apps and sources allow, as you.
  • Full test (Run a full test) runs the saved draft with its selected model and sources, if any, the way a real run works: it can use its read tools within the agent’s own model-call budget. It reads the agent’s attached Library files whole, however large; a real run opens large knowledge one file at a time. An agent with only a prompt needs no connected account. Tests never write to a connected account and can be stopped. The whole test is bounded to five minutes, including admission and result checks; its execution is bounded to three minutes or a lower saved run budget. A test with per-run sources starts only when none of your other agent runs is in progress and the agent has fewer than two runs going, and not after you have started 20 agent runs and tests in the last hour; Hive tells you which applies, so you can try again.
  • A test of an agent that needs no per-run sources, connected accounts or attached knowledge runs as the agent’s own read-only run, which is usually much faster. It answers on the agent’s model with the agent’s instructions, searches the web only when the agent allows it (a search may use what the run has read), and returns the agent’s JSON output when it declares one. It reads the apps attached under Integrations and the sources the agent may read (such as your connected systems or documents) as you, exactly as a real run would — so a test of an Outlook agent searches your own Outlook mailbox. It never carries out an action, even when the agent has actions selected: when the agent would use one, the Test shows it under Would do (“Would: create a Gmail draft to …”) and marks it Not carried out. When your test message asks for a chart, the Test draws it from the agent’s attached spreadsheets, a spreadsheet it reads or numbers you typed, and shows it under the answer with its source. A draft preview of an agent with attached knowledge also runs this way, reading that knowledge and the draft’s attached apps and sources. Every other test works as described above, so a test with per-run sources keeps its versioned source receipts. A test and a real run (from the agent’s run page, a run form, a schedule or a workflow step) treat two things the same way: what you type into the agent’s form counts as your own words, so the agent can search the web with it and open a link you entered; and a read of an attached app that has no typed read picks and summarises its records on the agent’s own model, not the workspace default.
The Test shows its work. While it runs, Hive is working lists each step as it happens: what it searched or read, and the actions it would take. When the answer lands you see:
  • Worked (or Worked, thought when the model summarised its reasoning): every step, with how many sources each read returned.
  • Sources read: the files, records and pages the reads returned, with links where they have one.
  • Tools called: each tool the agent used and what it found.
  • A summary line: the model, input and output tokens, what the Test cost in credits, shown in pounds at 1p a credit (for example “£0.03”, “Under £0.01”, or “Paid with your own model key” when your organisation’s own key is in use), and how long it took. On Business, a test and an evaluation use your own AI credits. On an earlier plan, a test that stays inside the preview limits (the agent allows at most 5 model calls, and no call reads more than 100,000 tokens or writes more than 8,000) draws first on the workspace’s test credits (200 a month for each Pro or Max seat, not carried over), then on the shared pool; a larger test, and an evaluation, draws on the pool.
Steps, cost and duration are shown for tests that run as the agent’s own read-only run. Other tests show their answer, tools and tokens. Start from a conversation starter or your own question.A test never stands in for a real run. To run the deployed agent and see its saved runs, select Run this agent and see its runs in the Test pane: it opens the agent’s run page.
8

Save and deploy

Step 4, Review & deploy, checks the agent before it goes live. Four chips show whether it is Named, Has instructions, Has a way to start and Has knowledge or tools; select one to open the step that fixes it. A card per step sums up what you set, each with Edit. Under them, Hive says whether a workflow can use the agent (Usable in workflows, or what to remove), and what a run of the saved version costs at most.The deploy action is the same in the header, on the last step’s Next and at the bottom of Review & deploy:
  • Deploy (or Resume for a paused agent) saves your changes first, then asks you to confirm. It needs a name, instructions and, for a form, a name for every field. Knowledge or tools are advice only: an agent that answers from its instructions alone can go live.
  • Deploy also checks the agent’s model. If your workspace can’t use it, Hive names the model, says why and what to do, for example “This agent’s model, GPT-5.5, is not available in this workspace: …”. Choose another model under Model in Basics & instructions and deploy again. An Owner or Admin is told the exact reason: an admin turned the model off in Settings → Models, its provider has retired it, or Hive can’t run it right now. Everyone else is told the model and the next step, and to ask a workspace admin to check Settings → Models. See A model is missing or unavailable.
  • Save draft saves a draft without deploying it. It shows once you have changes.
  • For a live agent, Publish changes saves your changes, and the new revision is live straight away for direct runs and Ask. When nothing has changed it reads Update and says everyone already has the latest version.
Workflow steps stay pinned to the revision you selected: choose the newer revision in the workflow builder and deploy that workflow to update its runs. Version history in the agent menu lists every saved revision.

Templates by Hive

Browse templates opens the template library. Every template in it is made by Hive and labelled By Hive. Choose a category, or search by name, integration or tag, then open a template to see everything it brings:
  • its input type, Prompt or Form, marked This template. The cards show the template’s choice and can’t be changed here; switch it in the editor after you configure the template. Expand What this does on a card to see what the agent Reads (its input, your documents or connected apps, knowledge files, the web and its read actions) and what it Changes (its available write actions). When a template lists more automatic changes than one run may make, it says Makes at most 3 changes per run; a template with no available write action says Takes no external action;
  • its form fields, with each field’s type and whether it’s optional;
  • its full instructions and conversation starters;
  • its actions, grouped by the connector they use. Select an action to see its parameters, with the required ones marked *. Each action is marked Available or Not yet available in Hive, and a connector Hive doesn’t offer yet is marked as such;
  • its long description, integrations, suggested knowledge files and tags.
Configure template creates a draft agent you own and opens it in the editor; nothing is live until you deploy it. The draft carries the template’s description, instructions, form fields, conversation starters and tags, and any file template it writes its documents on. A template never connects an account for you. Its agent reads each system through the invoking member’s own connected account, as that member, and says so when a system it needs is not connected. Reads need no setup. A template’s apps come attached under Integrations, with its Available write actions selected and nothing else: for example Create draft message and Move message to folder in Outlook for Outlook Inbox Manager. Open an app in the editor to see them switched on. An action that sends something straight away, such as Send email in IT Help Desk or Google Meet Meeting Transcription, is available but left off: switch it on yourself if you want the agent to send. An agent with attached apps reads only those apps. Templates that only read attach their apps with no actions, so they can’t take any. Templates with actions carry out their selected actions when they run, as the runner’s account level allows (straight away on Act); one whose recipient, channel, number or link did not come from the agent’s instructions or the runner’s own words waits for approval. You can change the selection, or set What it may do to Only reads and answers, before deploying. An action marked Not yet available in Hive is never selected; when Hive adds it, new drafts from the template pick it up. A template’s file field, such as Attachments/Refs in IT Help Desk or Photo in Field Incident Logger, becomes a run source: whoever runs the agent can choose Library files for that run. Templates that work from knowledge, such as Legal Help Desk and Translator with Glossary, list the files their design suggests under Knowledge, but Hive can’t include them: attach your own contracts, voice guide or glossary as knowledge in the editor. Open Register Due Diligence reads UK companies from Companies House when your workspace has it switched on, in place of the German OpenRegister. Each template’s agent uses the model you choose in the editor. If that model is later retired by its provider, the agent keeps running on the next model of the same tier, and so does a Test of it in Agent Studio. The editor no longer offers the retired model, so choose a current one when you next edit the agent (see Models that are retired).

UK accounting templates

Under Browse templates, choose Accounting for five templates for UK accounting practices. Each is a Form template. None sends anything: their only write actions save an email as a draft in the runner’s own Outlook or Gmail, or prepare a signature request as a draft, for you to review and send yourself, and Engagement Letter Drafter, Fee Quote / Proposal and New Client Onboarding Pack save their letter or proposal as a new Word file in your Library when you run them (see Files a run writes to your Library). AML Evidence Pack has no write at all. Each reads Companies House by the client’s company number, from the form or from your Client register, and searches by name only when there is no number anywhere. These templates chain several tools, so an agent made from one runs on larger budgets than the default: 12 model calls and 200,000 input tokens, and 16 model calls and 250,000 input tokens for New Client Onboarding Pack. An agent made from one of these templates before 10 November 2026 keeps the budgets it was made with. Because the input budget is above the default, starting one of these agents with Run in Agent Studio asks you to confirm when Hive estimates the run above £5 (500 credits on an earlier plan), as any agent with a raised token budget does.
  • Engagement Letter Drafter asks for the client name or company number, entity type, services in scope, fee basis, year end and the signer’s name and email. It reads the company’s profile, officers and persons with significant control from Companies House and writes the letter on Hive’s Engagement letter starter, set as the agent’s Word file template, as a new file in your Library. When the services include tax work, it reminds you that the client must authorise your firm with HMRC, linking HMRC’s client authorisation guidance and form 64-8. Choose your price list as the optional Price list for a run, and the letter’s schedule of fees comes from Hive’s fee calculator rather than the model. It can save a draft email to the signer, and when you ask, it prepares the letter for signature, sent after your approval: a draft signature request with the signature fields placed, which you review and send yourself. See Engagement letters for accounting practices.
  • Statutory Deadline Chaser asks for the Run date, because an agent is not told today’s date, and reads your Client register, attached as knowledge or chosen as the Client register for a run. An Engager client export works as it is: the agent reads its columns by the names on the starter’s Engager columns sheet. It reads each company’s accounts and confirmation statement due dates and filing history from Companies House, and its officers’ and PSCs’ identity verification: a director or PSC with a verification statement due is listed on that date, and a director with nothing on the register on the company’s next confirmation statement date, when an existing director must give it. Where you have connected HMRC, it also reads each client’s open VAT return periods (for a client with a VAT number) and Making Tax Digital Income Tax quarterly deadlines (for a client with an NI number), labelled HMRC. Other HMRC dates are worked out from your register by HMRC’s usual rules and labelled as estimated: a VAT return one month and 7 days after each quarter of the client’s VAT stagger (when HMRC could not give it), corporation tax 9 months and 1 day after the Year end and the CT600 12 months after it, and Self Assessment on 31 January (and 31 July, for payments on account) for a client with a UTR who is not a company. From the register’s gross income columns it puts each sole trader or landlord in the April 2026, 2027 or 2028 Making Tax Digital for Income Tax cohort, as an estimate, and lists the quarterly updates and final declaration from the tax year the client joins; with HMRC connected it reads HMRC’s own MTD Income Tax status and final declaration deadlines first (see Making Tax Digital for Income Tax cohorts). When HMRC is not connected or refuses for a client, it says so and links HMRC’s client authorisation guidance and form 64-8. It lists anything overdue or due within 60 days and suggests reminder text. It can save up to three draft reminders in one run, for the clients due soonest, and gives the rest as reminder text, naming the clients that have no draft.
  • AML Evidence Pack takes the company and your firm’s risk questionnaire answers, and assembles a client due diligence (CDD) evidence pack from Companies House (profile, officers, persons with significant control, disqualified directors, insolvency and charges), The Gazette and the FCA Register. It is an evidence pack, not an AML check: it performs no ID&V or sanctions screening, gives no risk rating and Takes no external action. Your firm’s own checks and supervised provider stay responsible for those.
  • Fee Quote / Proposal takes a scope form (services, turnover, transactions, record quality, start date, discount and whether to charge VAT) and prices it from the price list you attach as knowledge. Hive’s fee calculator works out every fee, the annual and monthly totals, any catch-up, the discount and VAT, and shows its working; the model never does the arithmetic. It then drafts the proposal, with that schedule of fees, on Hive’s Client proposal starter, as a new file in your Library. It never invents a fee: a service missing from the list is left as a placeholder. It can save a draft email to the client, and prepare the proposal for the client to accept and sign, sent after your approval. Lay your price list out as the Price list starter is; see Fee quotes from your price list.
  • New Client Onboarding Pack asks for the client name or company number, entity type, the contact’s name and email, the services in scope and, optionally, the client’s figures for pricing and their tax references. It works from the price list and Client register you attach or choose for the run, and prepares in one answer:
    • the Companies House profile, officers and persons with significant control;
    • the AML evidence pack (disqualified directors, insolvency, charges and Gazette notices) with a Risk questionnaire section. The member records the risk as Low, Medium or High: Hive never decides it, and performs no ID&V or sanctions screening;
    • the fee quote, worked out by Hive from your price list, never by the model. A price list you choose for one run is priced the same way as one you attach. If Hive cannot price a service, the agent says why, leaves marked placeholders in the schedule of fees and carries on with the letter and the other steps;
    • the engagement letter on the Engagement letter starter, saved as a new file in your Library and, when you ask, a signature request prepared as a draft for you to check and send. The draft shows as a card under the run’s answer, addressed only to the contact email you gave in the form (or one you name in your message), and nothing goes to the client until you open it and press Send;
    • HMRC authorisation: with HMRC connected with your firm’s agent reference number (ARN), it checks straight after the Companies House reads whether the client has already authorised your firm for MTD VAT or MTD Income Tax, and gives the exact request to make for each one still missing. Without that connection it gives the authorisation guidance and form 64-8 instead. It never makes the request: you make it in Ask, where it waits on an approval card when your HMRC account is set to Draft. For Self Assessment, PAYE, CIS and corporation tax, which HMRC’s API does not cover, it links HMRC’s client authorisation guidance and form 64-8;
    • a records request checklist email to the contact, saved as a draft when you ask;
    • a ready-to-paste row for your Client register, using the register’s own drop-down values even when your register is an Engager export (see Moving from Engager).
    It sends nothing.
These templates read Companies House, The Gazette and the FCA Register only while your workspace has each register switched on (see Switch it off for your workspace). If one is off, the agent says which and works from what you paste. Change the file template under File templates, in the editor’s Knowledge & tools step. Statutory Deadline Chaser and New Client Onboarding Pack read HMRC through the runner’s own HMRC (Making Tax Digital) connection, attached under Integrations with no actions. HMRC is a preview offered in some workspaces only, and answers only while the runner is using Hive in a browser, so a scheduled run says it could not read HMRC. A template cannot select an HMRC action for you. The fee calculator and signature requests come from Hive’s own tools: if either is not available in your workspace yet, the agent says so and leaves that step to you.

Investor templates

Two Form templates are written for venture capital and private equity investors. Both are read only: they read your own Gmail and Google Drive as you, and never send, draft or change anything. Configure template makes your own agent from one, which reads with your own connected accounts. Hive suggests them automatically once you connect those apps, in investor workspaces where Hive has turned that on.
  • Pitch intake asks which pitch email to read (a sender, company or subject words) and, if you have it, a Google Drive link to the deck or a deck from your Library. It reads the email thread and the deck and writes a one-page intake note: the company, the ask, traction, team, risks and questions for the founders. It writes not stated rather than guess, and Deck not read with how to fix it when it cannot open the deck.
  • IC memo builder asks for the company name and a Google Drive folder link, or Library files you pick, and drafts an investment committee memo: summary, thesis, market, product, traction and financials, team, risks and valuation questions. Every figure names the file it came from, and it may run up to four web searches for market and competitor context, each cited. The memo is a draft for discussion and makes no recommendation.

Due diligence starters

Under Browse templates, choose Due Diligence to start from one of 20 evidence-review jobs, from scope and data-room coverage through company analysis, chapter reviews and report QA. These are reusable read-only starters, not pre-built client engagements or automatic report writers. Output Composer is not offered: rendering and approving a slide deck remain separate work. Each starter asks for a target or project name and requires you to choose a Project Library folder for each run. The runner can choose only a folder they can currently access. Add the relevant materials to Library first; an empty or incomplete folder cannot support a complete finding. The draft distinguishes facts, assumptions and missing evidence, and a consultant must review its conclusions before using them. No starter contacts a client, connects to a data room on its own or sends work to another system. For a repeatable team process, save and deploy the agent, then map its required folder in a workflow with the initiating member’s access. Each starter needs a Library folder for every run, which a public visitor cannot provide, so Get public link explains this instead of creating a link. Use the authenticated direct form or Ask for them, or make the folder optional before sharing a public link.

Give an agent a project

An Ask project’s files are a Library folder, so an agent can take a project as a source. In the agent’s knowledge, choose + Add folders and pick the project’s folder. Project folders are listed as Project files, and searching for “project” finds them. The agent reads the project’s files as they are when it runs, with the runner’s access: only the project’s people can read them through the agent. When you save a chat that belongs to a project as an agent (see Save a chat as an agent), the agent takes the project as a source by itself, with the project’s instructions as its guidance for those files.

Choose what it returns

In step 3, under What it returns, pick Text or Structured data. Text is a written answer, with headings, lists and tables when they help; tick Plain text only for a system that can’t show formatting. Structured data gives a workflow named values instead of prose, using the same schema as a workflow Agent step’s structured output. Add named fields under Output fields, or write the schema under Advanced output schema, where an object may also name no fields or allow other keys ("additionalProperties": true). It works on every model the agent can use: Hive checks each answer against the schema before it is kept, and an answer that does not fit goes back to the model to fix while the run has model calls left.

Use it in a workflow

Every deployed agent is available to workflows: save the agent and make sure it is live. In the workflow builder, add the agent from the Agent library heading in Add step, or open Use a saved agent instead on a Hive agent step and choose it, then map its required inputs. Each mapping saves as you make it. Inputs whose key matches an earlier value are mapped when you choose the agent, and any input still unmapped suggests its match. For a Library file or folder source, choose Enter a fixed value and select an accessible file or folder by name under Source mappings; you do not need to paste an internal resource ID. Then deploy the workflow. The step keeps its selected agent revision, and each run checks the initiating member’s access to the mapped Library source again when the step starts. While the step runs, cancelling the workflow run, pausing or archiving the agent, removing the person who started the run (or moving them below Operator) or turning on the kill switch stops it before it acts, reaches outside Hive or saves its result. When a run continues after a Wait or an approval, it uses each finished agent step’s saved result as it is: the run fails if the agent is no longer live or the person who started the run has been removed or moved below Operator, but the step’s sources are not checked again. Public workflow forms cannot run a saved agent.

Share the agent

Use Share to give named collaborators access. They can use the agent with their own connected accounts; sharing never hands over yours. Admins can see workspace agents but can’t take over your credentials.

Other things you can do

The agent menu also offers:
  • Pin
  • Pause
  • Duplicate
  • Export agent, which gives you a portable file that Import agent can bring back in as a draft
  • Configure delegations (Business, Team and Enterprise), which lets the agent hand sub-tasks to up to 20 other agents, up to 5 levels deep. On Starter the dialog says delegation is not included in the plan instead of offering Add delegation; a delegation added earlier can still be switched off or removed. The API and AI clients connected over MCP follow the same rule: on Starter, saving an agent with a new delegation, or switching an existing one back on, is refused with feature_not_in_plan. A switched-off delegation is kept as it is when you duplicate the agent, import it or restore an earlier version: switch it off first if you need a copy on Starter. An agent used in Ask or run from Studio can ask the agents listed under Sub-agents itself, while it works: each one runs as the same member, within the budgets of both agents and the delegation, at most two levels deep and never back to an agent already working on the request. Its answer and sources come back to the first agent. It takes only the actions selected on its own apps, within the external writes the run has left (a new delegation allows none until you raise External writes under Delegation budget ceilings), and each one is recorded in the same chat or run as the first agent’s. Write where a sub-agent sends into the sub-agent’s own instructions: the first agent’s instructions reach it only as the request the first agent writes, which does not count as words you trust.
  • Activity timeline
  • Version history
  • Archive
  • Delete permanently

Writing good instructions

“Draft a polite reminder for each invoice more than 14 days overdue, one email per client, and list the invoices you skipped” works better than “handle overdue invoices”.
Name the Library folder or template to follow (“Use the reminder wording in Credit control templates”), and attach it as knowledge.
An agent with many tools starts each run with its most-used ones and finds the rest as it works. A tool the instructions name by its tool name, such as “write the letter with create_library_file” or “price it with calculate_fee_quote”, is ready from the agent’s first step instead, in a run from its run page and in a chat with it (up to 16 named tools). Hive’s fee calculator is ready before the agent has opened your price list, so a step that prices with it never waits for one. An agent with a Word or Excel file template always has the tool that writes its file ready on its run page. Naming a tool never gives the agent a tool it does not already have.
List hard limits such as “never offer a discount” or “never email a client marked on hold”. These guide the model. The policy gate enforces actual permissions whatever the instructions say.
Write every address, channel and link the agent may send to into its instructions: “email the weekly figures to finance@yourfirm.com”, “post alerts to #ops-alerts”. An action to a place the instructions or your own words name runs by itself; one to a place the agent found in what it read, such as an address in an email, waits for your approval. See the tip under Connect its tools for how to name channels, chats and repositories.
On Act, a selected action runs without a review unless where it sends came from something other than the instructions or your own words, so start with actions that are easy to undo — Create draft rather than Send email — and a write budget of one. Add more after a track record. Test it with Draft preview, which never touches your tools. See What agents and workflows do by themselves.

Publish and evaluate

Keep quality high with analytics, feedback and schedules.

Agent Studio overview

Statuses, roles, safety and limits.