Agents

What an agent is

An agent is an AI employee with one clear job, such as answering customer questions or qualifying new leads. You write its job description, give it the facts it may use, choose what it may do, and decide when it must ask you first.

Agents belong to a project. Open Agents in the sidebar to see the agents in your current project. See Find and organise agents.

Create an agent

You need the Owner, Admin or Builder role.

1. Open Agents and select New agent. In an empty project, you can also pick Support desk, Lead qualifier or Start blank straight from the page.
2. Choose a template: Support desk, Lead qualifier or Blank agent.
3. Give the agent a Name of at least 2 characters, and select Create agent. The agent opens in the builder as a Draft.

To start from an agent someone exported, select Import a file in the same dialog, or Import at the top of the Agents page. See Export and import.

The templates are starting points, and you can change everything before you publish:

• Support desk answers customer questions from your FAQ in the customer's language, and hands refunds, complaints and anything it can't answer to your team. Replies ask for approval first, and then go out by themselves after 2 hours if nobody decides.
• Lead qualifier asks new enquiries one question at a time, saves each lead, checks it against what makes a good lead for you (with a score from 1 to 5 worked out from your criteria), and tells your team.
• Blank agent starts with only the basics switched on.

Every template except Blank agent comes with an example business we invented, marked "(example)": notes, a price or slot table where one fits, and ready tests that pass on it. Your new agent starts without it, so nothing invented mixes with your real information. To try the template straight away, tick Start with example data when you create the agent: the notes go into its Knowledge box, and the table becomes a knowledge table named "Example: …" attached to it. It is not your business, and we will not let you publish it by accident: an agent that still has it is marked Our example text in the agents list and in the builder, and Publish agent stops and shows you a line of it first. You can still publish it while you are testing — you just have to say so, and it goes in the workspace audit log.

Each template's page on our website shows who it suits, what it does step by step, example conversations in English and Hinglish, the questions it asks leads, its ready tests and an estimate of a month's cost. Templates

The quickest way past it is to give the agent your own information instead. See Set up your first agent.

What you set

The Build tab has these settings. Changes save as you type; the header shows Saving… and then Saved.

• Description: one line for your team, shown in the agents list. Up to 300 characters.

• Instructions: the job description in plain sentences. Say who the agent helps, what it should and should not do, the tone to use and when to hand over to your team. Up to 8,000 characters. Use Insert variable to add a {{variable}}. See Variables.

• Variables: details you use in more than one place, such as your shop name or helpline.

• Knowledge: the facts the agent may answer from, pasted as text. Up to 20,000 characters. See Knowledge.

• Knowledge files: files and notes from the project's Knowledge page that the agent searches. Up to 20 per agent.

• Tools: what the agent may do besides replying, and whether each tool needs your approval. See Tools.

• Apps: actions in apps your workspace has connected. This is a pilot. See Channels and connected apps.

• How smart: Smart (Claude Haiku 4.5, the default: quick, low-cost answers for FAQs, bookings and leads), Smarter (Claude Sonnet 5: better with long conversations and tricky questions, about twice the cost) or Smartest (Claude Opus 5: the most capable, about five times the cost). The builder shows about what a reply costs on each, and every test reply shows its real cost.

• Welcome message: the first message people see when they start a chat. Say that they are talking to an AI assistant. Up to 500 characters. Variables work here too.

• Long-term memory: notes the agent keeps between conversations. Off by default. See Long-term memory.

• Task fields: details the agent records on each task, such as an order number. See Task fields.

• Let people send files: whether people can add photos and PDFs to their messages in the test chat, Team chat, email and the API. On by default. Website chat has its own switch, off by default. See Chat with your agents.

• Advanced settings: a folded section with the backup model, reply length and language, suggested prompts, a usage guide, hiding contact details, and these two limits:

  • Autonomy limit: the most steps (model calls and tool uses) the agent may take on one message, from 1 to 50. At the limit it stops and hands the task to your team, so a confused agent can't run up costs.
  • Auto-approve after: for tools set to ask and then run, how many hours to wait before the action goes ahead by itself. From 1 to 72 hours; the default is 2.

  See Advanced settings.

Variables

A variable is a named detail, such as your shop name, helpline or opening hours. Add it once in the Variables section, then write {{key}} in the instructions or the welcome message. When the agent runs, it sees the value in that place. Change the value, and every place that uses it changes too.

1. In Variables, select Add variable.
2. Give it a Name (for your team), a Key (small letters, numbers and _, starting with a letter, up to 40 characters) and a Type: Text, Long text, Number, Web address, Email or Phone.
3. Enter the Value, and select Add variable. The value must match the type; for example, a web address starts with https://.
4. In the instructions, put the cursor where the value should go and choose Insert variable.

Good to know:

• An agent can have up to 30 variables. Text values can be up to 200 characters, and long text up to 2,000.
• Placeholders are highlighted in the editor. If a {{key}} has no variable, it is highlighted in amber, a warning lists it, and the agent sees it exactly as written.
• Values are filled in once and shown to the AI as plain text. A value that itself contains {{...}} is not filled again.
• The test chat shows the welcome message with its values filled in.
• Don't put passwords, bank details or anything secret in a variable.

Long-term memory

With long-term memory on, the agent keeps short notes between conversations, like "Ramesh ji prefers delivery after 6 pm" or "We are closed on 2 October". Each time it replies, the notes that best match the conversation (and then the newest ones) are added to what it reads, marked as information and not as instructions.

• Turn it on in the Long-term memory section. It is off for new agents.
• Turning it on adds the Remember for later tool, which lets the agent suggest new notes. It is set to Ask first, because a customer could try to plant something false; each suggested note waits in Tasks, and you can correct it before you approve. You can change this under Tools, or untick the tool so the agent only reads notes.
• You can also add, search, edit and delete notes yourself. Each note shows where it came from: added by a teammate, or saved by the agent in a conversation, with a link to that task. Notes from test chats are kept too, and marked Test chat, so delete them if you don't want them used.
• Each note is up to 500 characters, and an agent keeps up to 200 notes. The same note is never saved twice.
• While memory is off, the agent doesn't read or save notes, but the notes you already have stay.
• Deleting the agent deletes its notes.

Task fields

Task fields are details the agent fills in while it works on a task, such as an order number, a size or whether the person wants a call back.

1. In Task fields, select Add field.
2. Give it a Name and a Key, and choose a Type: Text, Number, Yes or no, or Choice. For a choice, list 2 to 20 options, one per line.
3. Select Add field.

While an agent has fields, it gets the Record task details tool, which runs without asking. The agent records a value only when the person gives it, and the value must fit the type: a number for Number fields, yes or no, or one of your choices. The values show in a Task fields card on the task page, and as extra columns (named "Field: …") when you export Tasks. An agent can have up to 10 fields. If you delete a field, values already recorded stay on past tasks.

Other agents it can ask

An agent can ask your other agents a question and use the answer. Your support agent can ask your order agent "has order 1182 shipped?", without the customer being handed over.

1. On the Build tab, find Other agents it can ask and tick up to 5 agents.
2. Choose how each runs, like any tool: Runs freely asks straight away; Ask first waits for your approval, then the answer is added to the conversation when it comes.
3. Publish the agent.

The agent sees each one as a tool, "Ask Ravi Orders". The other agent can't see the conversation, so the asking agent sends every detail it needs. Its answer, and any task fields it records, come back to the asking agent.

• Its own task. The agent asked works on a task of its own, marked Asked by an agent and linked both ways: the asking task lists "Asked other agents", and the other one shows Asked by.
• Limits. Agents can ask each other two levels deep at most, and never in a circle. What they spend counts towards the first task's spending limit.
• Same rules. The agent asked uses its published version (its draft in a test chat), its own tools and approval settings. A paused agent can't be asked.
• Workspace rules can make Asking other agents always ask first, or switch it off. See Rules for every agent.

Advanced settings

Open Advanced settings near the end of the Build tab. It opens by itself when something in it was changed.

• Backup model: if the main model's service is down, overloaded or returns an error, the agent tries that reply once more on the backup model, which then answers the rest of that message. The step shows Answered by the backup model in the task, and it costs what the backup model costs. Spending limits and the platform pause are never worked around: when one of them stops a reply, nothing is retried. Without a backup, the task fails and you can run it again.
• Reply length: Short (one to three sentences), Medium (the default) or Long. It also caps how much text one model call can produce, so a very long answer is cut off at the limit.
• Reply language: Same as the customer (the default), or always English, Hindi or Hinglish.
• Suggested prompts: up to 4 starter questions, each up to 80 characters. They show as buttons when a test chat starts.
• How to use this agent: a note of up to 1,000 characters for your teammates, shown at the top of the Test panel.
• Hide contact details from Viewers: hides phone numbers, email addresses and card numbers from teammates with the Viewer role in Tasks (the list, task details and the export), in the agent's leads and in its memory. It only changes what Viewers see; nothing stored changes, and Owners, Admins and Builders see everything. Switching it on works straight away. Switching it off takes effect when you publish.
• Hide ID and card numbers from the AI (on by default): Aadhaar, PAN, card and bank account numbers, OTPs, CVVs and UPI PINs that customers type are replaced before the AI reads the message, for example with [card ending 1111] or [Aadhaar ending 1234]. You still see the message as it was sent. See What the AI never sees.
• Autonomy limit and Auto-approve after, described above.

There is no creativity (temperature) setting: the Claude models behind Smarter and Smartest don't accept one.

How agents behave

Whatever your instructions say, every agent:

• says it is an AI in its first reply, and whenever someone asks;
• replies in the language the person writes in, such as English, Hindi or Hinglish, unless you choose a reply language in Advanced settings;
• answers only from its knowledge, tool results and what the person said, and hands over rather than guesses when it doesn't know;
• never asks for card numbers, UPI PINs, OTPs, passwords or bank logins;
• treats anything that comes back from a connected app as data, not as instructions.

Test it

The Test panel sits beside the builder. On a phone, switch to the Test tab.

• Write a message the way a customer would, or tap one of your suggested prompts. The agent answers from your saved draft.
• With Let people send files on, use the paperclip to add up to 3 photos or PDFs, such as a sample bill, and see how the agent reads them.
• If the agent has a usage guide, it shows at the top of the panel.
• Test mode: nothing is really sent. Replies and emails are simulated. In connected apps, actions that only read data run for real, and anything that would change data is simulated.
• Under each reply, What happened lists the steps the agent took and what the reply cost in rupees.
• Select New conversation to start again.

Each test conversation is also saved as a task marked Test, so you can open it in Tasks.

Publish, pause and resume

• Publish agent makes the current draft the live version. You can keep editing the draft; the live version changes only when you publish again. The header shows Unpublished changes when the draft differs.
• Versions lists every published version, with when it was published and by whom. Select Make live again on an earlier version to undo a change that went wrong. See Make an earlier version live again.
• Pause stops a live agent from taking real work. Resume makes it live again.

Customers reach a published agent through its channels, such as website chat, a share link and email. See Channels and connected apps.

Make an earlier version live again

If a change made the agent worse, you don't need to undo it by hand:

1. Open the agent's Versions tab.
2. Select Make live again on the version you want, and confirm.

Real conversations use that version's settings straight away. It is saved as a new version, so the list still shows what happened. Your draft is replaced with the same settings, so the next publish doesn't bring back what you undid.

Test sets aren't run again, because that version was live before. Everything it uses must still exist, such as its knowledge files, tools and connected accounts; if something was deleted, the app says what to fix. You need the Owner, Admin or Builder role. The activity log records who did it.

Leads

If the Save a lead tool is on, the agent has a Leads tab. It lists each lead with their need, budget, timeline, score and when it was saved. Leads from test chats are marked Test.

Find and organise agents

The Agents page lists every agent in the current project. For each agent it shows:

• its name and description;
• its status: Draft, Live or Paused, and Unpublished changes when the draft differs from the live version;
• how many tools it has: built-in tools plus app actions;
• how many knowledge files are attached;
• how many tasks it started in the last 7 days, and when its last task started. Test chats aren't counted;
• when its settings last changed.

Labels under the name show its folder, Website chat when the chat widget is switched on, and how many triggers are switched on.

To find an agent:

• Search by name or description.
• Filter by status, by folder, by Website chat on and by Has triggers. Select Clear filters to see everything again.
• Sort by last changed, by name, or by most tasks this week.
• Switch between the List and Grid views. Your browser remembers your choice. On a phone, the list shows as cards.

The page shows up to 500 agents at a time. If a project has more, search or filter to find the others.

Folders

Folders group a project's agents, for example by team or branch. They don't change how an agent works.

• Select Folders on the Agents page to create, rename or delete folders. A folder name can be used once in a project and can be up to 60 characters. A project can have up to 100 folders.
• To move an agent, open its ... menu and select Move to folder. You can pick a folder, create a new one, or choose No folder.
• Deleting a folder doesn't delete its agents. They move out of the folder.
• While you see all folders, the list is grouped by folder. Clear Group by folder to see one list. Pick a folder in the folder filter to see only its agents, or Not in a folder for the rest.

The agent menu

Each agent's ... menu on the Agents page has:

• Open;
• Rename;
• Duplicate, Move to folder, Copy to another project and Export (see below);
• Pause or Resume, for published agents;
• Delete, for Owners and Admins.

In the builder, the More actions menu has Duplicate, Export, Copy link to this agent, Move to folder and Delete. The link works only for people in your workspace.

Duplicate and copy

Duplicate makes a new draft in the same project and folder, named after the original with "(copy)" at the end. It has the same description, instructions, knowledge, knowledge files, websites, tools, app actions and settings.

Copy to another project makes a new draft in another project of the same workspace. The copy keeps the agent's name, and adds "(copy)" if that name is taken in the other project.

• Knowledge files and imported websites stay behind, because each project has its own library. Add them to the other project and attach them to the copy.
• App actions come along when their connection still exists. Connections belong to the workspace, so every project can use them. Actions whose connection was removed are left out.

A copy never includes the original's versions, triggers, website chat, share link, tasks or leads. After copying, the app lists anything it left out, and anything to fix before you publish, such as a connection to reconnect.

Export and import

Export downloads the agent's current draft as a .json file. The file has the name, description, instructions, knowledge text, tools and their approval settings, model, autonomy limit, auto-approve time and welcome message.

The file doesn't include knowledge files, imported websites, app actions, connection details or anything secret. It lists the names of the knowledge files, websites and app actions the agent used, so whoever imports it knows what to add again.

To import, select Import on the Agents page, or Import a file in the New agent dialog, and pick the file.

• The file can be up to 256 KB, and must be an agent file exported from Dhanur AI.
• The new agent is a Draft. Nothing runs until you publish it.
• Only settings are read from the file, and nothing in it runs. Text longer than the agent limits is refused.
• Tools this app doesn't have, app actions and knowledge file links are left out, and the app lists them.
• If a folder is selected in the folder filter when you import, the agent goes into that folder.

Imports, exports, copies, duplicates and folder changes are recorded in your workspace's audit log.

Delete an agent

Owners and Admins can delete an agent from its ... menu on the Agents page, or from the More actions menu in the builder. This deletes the agent, its versions, and all its tasks and approvals. Leads it saved are kept. It can't be undone.

Who can do what

• Owners, Admins and Builders can create, edit, test, publish, pause and resume agents. They can also manage folders, and duplicate, copy, export and import agents.
• Viewers can see, search and filter agents, their settings and their memory, but can't change, test or export them. If an agent hides contact details, Viewers see them hidden.
• Only Owners and Admins can delete an agent.

See Members and roles.

Later

Planned. An AI builder that drafts an agent from a one-line description of the task, more templates, and using your own AI model keys. To put several agents on one team, see Workforces.

Where agents fit

Each conversation with an agent becomes a task that you can follow step by step, with its cost in rupees. Actions you set to ask first wait for your approval. See Tasks and approvals.