Skip to main content
Once an agent is deployed, your job shifts from building it to keeping it good. Each agent has an Analytics tab to see how it’s used, a Feedback tab to read how people rate its answers, and schedules and public links to put it to work. Analytics and Feedback are for the agent’s creator and the people who can edit it. Members who can only use the agent don’t see them.

Analytics

Open the agent and select Analytics. Choose a window: Last 24 hours, Last 7 days, Last 30 days (the default), Last 90 days, Last 12 months, or Custom range with a start and end date (up to 366 days, in UTC, the end date included). The tab shows:
  • Usage: Active users, Conversations, Messages, Messages per user and Runs, with a chart of each over the window (by hour, day or month).
  • Adoption: how many people can use the agent, and how many of them used it in the window, by asking it in Ask or running it.
  • Conversations per user: how many conversations each active person had.
  • Cost: the settled model cost from credits, per message (answers in Ask) and per run (runs from the agent’s run page, scheduled and workflow runs, and exchanges on its public link), and in total. Missing costs are not counted as zero. Costs are counted from 12 October 2026, when Hive began taking payments: a window that starts earlier still counts its earlier messages and runs, but not what they cost.
  • Most used actions: the actions the agent carried out in its apps, by count.
  • Runs: recorded runs and how many succeeded.
Analytics is included on Business, Team and Enterprise. On Starter, the tab says agent analytics is not included in the plan, with a link to plans for Owners and Admins, and reading an agent’s analytics through the API or an AI client connected over MCP is refused with feature_not_in_plan; Feedback works on every plan. Analytics counts every member’s use of the agent, but never shows anyone’s conversation. Top users, which names the people who use the agent most, appears only when a workspace admin turns on Member-level analytics. It’s off by default. Admins see the switch on any agent’s Analytics tab, and it applies to every agent in the workspace. Select Export CSV to download the chart data, one row per hour, day or month.

Feedback

People rate an agent’s answers with a thumbs up or down, and can add a comment for the agent’s creators and the Hive team (the comment box says so):
  • In Ask, under any answer in a chat with the agent. After rating, a box offers an optional comment.
  • On a run, in the run’s details: Good run or Bad run, then an optional comment.
  • On a public link, the visitor can mark an answer Good response or Bad response with the thumbs under it. Visitors leave a rating only, with no comment and no name.
Open the agent and select Feedback to see the ratings in the window you choose: how many there are, the share that were positive and negative, and each rating with where it came from (Ask, Run or Public link), its date and its comment. Filter by Positive, Negative or With comments. Raters are named only when Member-level analytics is on. The answers that were rated always stay in their private conversations and runs. Select Export CSV to download the list. Rating again replaces the rater’s earlier rating of the same answer or run. Every rating and comment is also sent to the Hive team, so we can improve Hive: the rating, the comment, the workspace’s name, the rater’s role (or “visitor” on a public link), the agent and its model. The rater’s name and the answer that was rated are not sent.

Schedule an agent

Deployed agents can run on a recurring schedule, for example a Monday-morning pipeline summary or a daily overdue-invoice sweep.
  1. Save and deploy the agent. A schedule always runs a deployed revision.
  2. Select Schedule (the calendar button) in the agent editor. The Create scheduled task dialog opens; give the task a Name.
  3. Pick a Recurrence: Every day, Weekdays, Every week or Every month, and a Time. Weekly schedules use Days of the week, and monthly schedules use Day of month (months without that date are skipped).
  4. Set the Timezone (for example Europe/London) and the Instructions or inputs the agent should run with, then select Save.
Scheduling needs Operator access or above, plus access to both Agent Studio and the Scheduled product. A schedule pins the revision it was created with. If you later deploy a new revision, delete the schedule and create a new one for the new revision. A reusable agent with a run form (one that declares its own sources, context or output, such as an agent built from a reusable template) runs from Ask, its own page or a workflow step, so its Schedule button is greyed out with the reason “This agent runs from Ask, its own page or a workflow step. To run it on a schedule, add it to a scheduled workflow.” To run it on a schedule, build a workflow with a schedule trigger and an Agent library step for the agent. A scheduled run uses the AI credits of the schedule’s owner on Business (on an earlier plan, the workspace’s pool) and runs on an Efficient or Balanced model: an agent set to a Frontier model is refused when a schedule starts it, unless your organisation’s own model key is in use. If those credits have run out when a run is due, it waits up to 72 hours and starts once credits return (see Waiting for credits). To run a deployed agent once at a set date and time instead, ask for it in Ask (“run my Invoice chaser tomorrow at 9am”); see Build from chat. You can see and manage every schedule in Scheduled tasks. A public link lets anyone talk to an agent without a Hive account, for example a client intake assistant you link from your website or an email.
1

Deploy the agent

Only a deployed agent can have a public link. The agent’s owner creates and manages it, and must be a workspace admin or owner to create one.
2

Open Public link

In the agent editor, select Public link. This dialog is the only place that turns public use on.
3

Get and copy the link

Select Get public link, then Copy link. If the agent’s current version doesn’t allow public use yet, Hive turns it on first and saves a new version. Hive shows the full link only once, because it keeps only a fingerprint of it. If you lose it, select Get a new link. The old link then stops working.
There is no version to choose. The link always runs the agent’s latest deployed version. When you save or deploy a change, the link moves to it straight away, so visitors get it from their next message; you don’t need a new link. Someone who had the page open when you saved stays on it: the page opens the same link again by itself and starts a new conversation with the updated agent, with the note “This agent was updated, so a new conversation started.” If you revoked, paused or replaced the link instead, the page stops at once and says the link is not available.

What visitors can and can’t do

  • Visitors don’t sign in. Anyone with the link can use it, including people it’s forwarded to.
  • The page is a chat that looks like Ask, without Hive’s sidebar: the agent’s name and description, a message box, and the conversation. A line at the bottom says it’s an AI agent whose answers can be wrong; About this link explains that the agent is run by the workspace that shared the link, that anyone with the link can use it and that the conversation belongs to the visitor’s browser. Search engines are asked not to index the page.
  • Before their first message, visitors see the agent’s Conversation starters, from the version the link serves, under the message box. Selecting one puts it in the box, to edit or send. If the agent asks for details (form fields or a choice of sources), a Details chip in the message box opens them; an agent whose first turn is only a form shows its fields with Start instead, and no starters.
  • Each question and answer is its own turn in the conversation. While the agent works, its turn shows what it is doing and how long it has taken, and Stop response ends the request. Answers are formatted like answers in Hive. Under each, chips open the sources it used (by name, with the passage and when it was read), any files the agent produced (to download) and, for an agent with structured output, the Structured data; a chip shows only when there is something in it. Copy copies the answer. Earlier answers show their text, with Show sources and files to load the rest.
  • Anything that stops a request shows as a line in the conversation where it happened: a stopped request in grey; a limit, a question that needs too much at once or a link at its daily or usage limit in amber; a failed request in red, with Try again when trying again can help. A message that couldn’t be sent stays in the box with Try again. If Hive can’t be reached for a moment, the conversation stays on the page and a line above the message box says it is reconnecting. If Hive sends back an answer the page can’t verify, the answer and its files are taken off the page and a line says so, with Try again. Only a link that can’t be used at all replaces the page, with This link is not available.
  • New conversation at the top starts over once a conversation has finished. If the agent doesn’t take follow-up questions, the message box gives way to a line saying so, with New conversation, once the answer has finished; until then Stop response stays available.
  • Public runs give answers only. They never use connected accounts, take actions, or read Brain, other agents or anything else in the workspace. If a version you deploy needs any of those, for example a connected account, an action or a required Library folder for each run, visitors see that the link is unavailable until you deploy a version without it. The link itself stays live and picks up that version automatically.
  • An agent can answer visitors from the Library files and folders attached to it, for example a support agent with a folder of one customer’s documents. Each run reads only the files and folders attached to the version the link serves, with the access of the person who created the link. If that person leaves the workspace or loses access to a file or folder, the link stops answering until access is restored. Under each answer, visitors see the attached files the agent read for it: the ones it opened, or all of them when they are small enough to be read whole.
  • Up to about 50,000 tokens in total, each answer reads the attached files whole. Past that, the agent gets a list of them (name, date, first lines) and opens only what the question needs, within the exchange’s 100,000-token reading limit: each opening sends at most about 16,000 characters of one file (roughly 4,000 tokens), the passages that contain the question’s words when the file is longer, and one answer opens at most about 50,000 characters in all. Files are read as their text: a scanned PDF with no text can’t be read on a public link. A spreadsheet too large for one opening comes as a list of its sheets, each with its size and column names, and the agent then opens the sheet it needs by name, from any row, so a sheet at the end of a large workbook is read as easily as the first. In the same way, a PDF or slide deck too large for one opening comes as a list of its pages or slides, each with its number and first line, and the agent then opens exactly the page or slides it needs by number, so slide 24 of a long investor deck is read as easily as slide 1.
  • When the answer’s time or reading limit is about to run out, the agent answers from what it has read. A question that needs more than one answer can read, such as a summary of every attached document, gets This question needs too much at once: ask about one topic or document at a time. A file over 25 MB isn’t read at all. A public link never acts, searches the web or reads anyone’s connected apps.
  • With Library files attached, an answer can take up to three model calls, so the agent can total a spreadsheet exactly or draw a chart before it replies. Its only tools read the attached files. The link, its owner’s access and the kill switch are checked when the answer starts; the link and the kill switch again before the answer is saved and before it is shown.
  • A visitor who asks for a chart, such as “chart revenue by region”, gets one drawn from the agent’s attached spreadsheets or from numbers they typed, shown with the answer. The line under the chart names the file and sheet it came from, and nothing a visitor sees includes the workspace’s id for a file.
  • Each visitor’s conversation belongs to their browser session. Nobody else with the link can read it.
  • While the workspace kill switch is on, the link shows visitors that it is unavailable: it takes no new messages and shows no earlier answers until an Owner or Admin resumes.

Cost and limits

On Business, visitors’ exchanges use the AI credits of the member who published the link, like their own work, with no separate public budget; when the publisher’s credits and any extra usage they may use run out, the link stops answering until they have credits again. On an earlier plan, visitors’ exchanges use credits from your workspace’s shared pool, inside the public budget: at most 25% of the month’s included credits (your trial credits during a trial), for public agent links and public forms together. Rolled-over credits, top-ups and bonus credits don’t raise it. An Owner or Admin can lower it under Settings → Usage → Spend controls. Public traffic never triggers auto top-up, and it uses credit top-ups only if an Owner or Admin allows it there. When the public budget or the pool is used up, the link stops answering until it has credits again. With your organisation’s own model key, visitors’ model calls run on your key and each uses 1 credit from the public budget, never from the member who published the link. On an Anthropic key, visitors are served Claude Haiku wherever they would otherwise get Hive’s Efficient model. A link stops answering while the member who published it holds no seat: moving their seat to someone else, or a plan change that suspends their seat over the limit, pauses their links, and the Owners and Admins are told. Resume a link once its publisher holds a Basic, Pro or Max seat again. Each answer stays inside the agent’s own per-run limits and inside the link’s own: one exchange runs for at most 300 seconds and reads at most 100,000 tokens and writes at most 8,000, with no actions.

Model for visitors

Model for visitors in Public link decides which tier of model answers visitors:
  • Efficient, the default: less than a penny an exchange on Business (on an earlier plan, about 2.4 credits, so £24 covers about 1,000 exchanges).
  • Balanced, if you turn it on for a link that needs stronger answers: about 3p an exchange on Business (about 21 credits on an earlier plan). Visitors get Balanced only when the agent’s own model is a Balanced model. An agent on an Efficient model keeps answering on that model, and an agent on a Frontier, custom or -pro model answers visitors on Efficient.
Frontier models never answer on a public link. See Model tiers and credits. On Business there is no limit on how many links the workspace can publish. An earlier plan sets how many it can publish at once: 1 on Starter and 20 on Team. Publishing another at the limit is refused; Revoke link on one first (a paused link still counts toward the limit), or upgrade. During a trial, a workspace’s public links and public forms answer 100 times in total, and 100 more for each further week if Hive extends the trial. After that, a link tells its visitors that it has reached its usage limit and takes no more questions, and the workspace’s Owners and Admins are told once, in the Inbox under Needs you. Choosing a plan opens the link again. To stop abuse, every link is limited: Days are counted in UTC. Each visitor is counted by their own internet address, so visitors on different networks never share an allowance. One connection can also use at most half of a link’s exchanges in an hour, however it presents itself, so a single source can’t use up a link for its other visitors. What a visitor sees depends on the limit they reached:
  • This agent has reached a limit (wait a moment before trying again): the link’s hour, or all of its simultaneous exchanges are busy. The visitor can try again shortly.
  • You have reached today’s question limit (try again tomorrow, or ask the person who shared it): this visitor’s own 20 a day. Other visitors can still ask.
  • This link has reached its limit for today (try again tomorrow, or ask the person who shared it): the link’s daily cap, shared by every visitor.
  • This link has reached its usage limit: the link’s monthly cap, or a trial’s public answers. The visitor asks the person who shared it.
Only exchanges the link answers count toward these limits: a message refused because the link was busy or at its limit does not use up any of the visitor’s 20 a day. The live updates limit only matters to a visitor with the page open in more than three tabs: the extra tabs check for the answer every 30 seconds instead. A long run’s live updates renew on their own, without the visitor seeing it. An Owner or Admin can raise a link’s daily cap up to 4,800 exchanges and its monthly cap up to 31 such days under Exchange caps in Public link. Leave a cap blank to go back to the default. The 200-an-hour limit is never raised, and every exchange still draws its credits from the public budget, so a raise never lets public traffic spend more than that budget.

No question limits

For a link that many people must be able to use freely, such as an investor data room during due diligence, an Owner or Admin can tick No question limits (investor / due-diligence links) under Exchange caps and save. It is off by default, and each change is recorded in the audit log. With it on:
  • visitors and the link have no hourly or daily question limit;
  • the link answers up to 10 visitors at once (still 1 at a time per visitor);
  • each answer keeps its own limits on length and time;
  • the monthly cap still applies, so raise it too if you expect heavy use;
  • the automatic pause on an extreme spike, described below, still applies, and Owners and Admins are told if it pauses the link;
  • Hive’s protection against automated abuse still applies, with room for many visitors at once;
  • every exchange still draws its credits from the public budget, billed to the publisher as before.
If a link’s traffic in an hour reaches 10 times its usual hourly level, and at least 100 exchanges, Hive pauses it for review, records it in the audit log and tells the link’s publisher and the workspace’s Owners and Admins; Hive also checks every link once an hour for this. If Hive cannot pause the link, the notice says so and asks them to pause it in Public link. A new or quiet link is never paused below 100 exchanges in an hour, so its first visitors and your own testing do not pause it. Resume it from Public link once you have checked it. Pause or revoke a link whenever you need to stop it. Open Public link to see whether the link is active and which version it serves. Select Pause link to stop new messages for a while, Resume link to restart it on the latest version, or Revoke link to end it for everyone. A revoked link never works again. If Public link says public links can’t be managed for this agent, check that you own the agent and that it’s deployed with public use allowed. Creating a link also needs a workspace admin or owner. Then reload the page. When a visitor starts the public agent, the page follows that run through to its completed result and keeps earlier responses available within the same conversation. If the agent allows follow-up messages, the visitor can continue from that page without signing in. Generated files appear with the response and are opened through the visitor’s current public session; copying an artifact URL does not turn it into a separate public link. If a run is interrupted, the page checks its recorded status before allowing another submission, so retrying does not silently create a duplicate run. Pausing or revoking the public link stops new messages and prevents its visitor session from reading later runs. If Hive can’t be reached for a moment, for example while an update is being released, the conversation and any answer already on the page stay there. The page says it’s reconnecting and tries again by itself, and the visitor can select Try again to check sooner. Someone who opens the link during that moment sees that the agent is temporarily unavailable, with Try again.

Keep improving

1

Watch feedback and analytics

Look for negative ratings, their comments, and unusual cost per message or run.
2

Reproduce the problem in Test

Ask the failing question in the agent’s Test pane and read how Hive worked.
3

Edit, test again, then update

Change the instructions, knowledge or model, test again, and update only when the answer holds up.

Build an agent

Create, test and deploy an agent.

Scheduled tasks

Manage every recurring agent run in one place.