Browse documentation
Creating an agent
How to build a custom agent from one sentence or a template, answer the builder's questions, review the draft, and provision it paused so nothing runs until you enable it.
One sentence is enough to start an agent
You do not write a custom agent from scratch. You say what the job is, the agent builder asks a few questions, and it drafts the whole agent for you to review: its skills, tools, triggers and guardrails included. Nothing runs until you have read the draft and switched the agent on.
There are three places to start:
| Where | What you press | Best for |
|---|---|---|
| Agent Workbench | + New agent | Building one agent step by step |
| Build AI Team | The chat at the top | Saying what you want and letting Nagent pick the builder |
| A new team's form | Build agent | A team that needs an agent you do not have yet |
Build an agent in three steps
Press + New agent in the Agent Workbench. The screen is titled Build an agent, and a strip at the top shows where you are: 1. intent, 2. questions, 3. review.
Step 1. Say what the agent does
Under Describe the agent in one line, write the job in plain words. For Prompt World, which runs prompt engineering workshops, that might be:
Watch the five competitor workshop pages we track every week and tell me when one changes its price or its dates.
Press Continue.
Or pick a card under Or start from a template, such as Competitor pricing watcher or Outbound follow-up nudge. A template skips the questions and takes you straight to the review, already drafted.
Step 2. Answer the questions
The builder asks a few multiple-choice questions about your agent. They usually cover how it should run (on a schedule, on an event or only when you start it), how much it may do on its own, a daily spending cap, whether it must ask before changing anything, and where its results should go.
Every question has an answer already chosen, marked recommended. Accept the defaults or change them, then press Build agent. The screen says "Assembling the agent" while the builder drafts it.
Step 3. Review and provision
The review has two halves.
- On the right, the Spec preview. The draft agent, field by field, and you can edit any of it by hand.
- On the left, the Agent builder chat. Ask for a change in plain words, for example "make it run daily instead", and the builder redrafts.
| Field | What to check |
|---|---|
| Key and Name | The key is the agent's address: lowercase letters, numbers and hyphens |
| Description, Capabilities | That it says what you meant |
| Tools | The actions it may use. Add or remove them here |
| Skills | The procedures it follows. The builder may write new ones when you provision |
| Triggers | What starts it: a schedule, an event or a manual start |
| Guardrails | Each rule, whether it is advisory or blocking, and what it does when broken: warn, queue for approval, block, or pause the agent |
| Model | Leave the default unless you have a reason |
| Autonomy | Its starting level |
| Risk level | Low, medium or high |
| System prompt | The agent's standing instructions |
The Autonomy field numbers the five levels L0 to L4, the way an agent's Autonomy pane does: L0 here is L1 Observer everywhere else. See autonomy and governance.
If something in the draft is wrong, a Validation issues box lists it. An issue marked as an error keeps Provision agent unavailable until you fix it.
When the draft is right, press Provision agent.
After provisioning
The agent is created switched off, and the left half says Provisioned (paused). It also lists:
- Auto-created skills, if the builder wrote new ones for this agent.
- Not usable yet, if a tool the agent needs is not ready. Each line says either that the app is not connected yet, or that the action needs ticking on Integrations. See tools and integrations.
Press Enable now to switch the agent on, or Open workbench to go to its page and keep configuring it first.
Our logic
An agent that starts paused cannot surprise you: you read the draft, check its tools and caps, and choose the moment it starts. And an agent that is missing a tool says so on the spot, rather than sitting switched on and idle with nothing to tell you why.
Asking from the Build AI Team chat
Build AI Team opens with one chat for everything that makes an agent. Describe what you want built and Nagent works out whether it is an agent, a skill or a workflow, restates it, and asks the right builder. If it cannot tell, it asks you to pick agent, skill or workflow.
To skip that step, start your message with a command. Type / to see them.
| Command | Asks |
|---|---|
/agent | The agent builder, for a teammate with a role that runs on its own |
/skill | The skill builder, for a reusable instruction an agent follows |
/workflow | The workflow builder, for a sequence that hands work between agents |
/build | Nobody yet: describe it and let the chat decide |
A builder asked from the chat creates nothing on its own. It queues a proposal and the chat says so, with an Open link. Approve it on Actions, under Awaiting approval.
Building an agent while forming a team
The New team form lists the agents in your workspace. If none of them fits, write the job under What it would do, in one line and press Build agent: the builder opens with your sentence. Come back to the form when the agent exists, and it is on the list. See setting up a team.
Where to go next
- The Agent Workbench: the agent's own page, where you keep configuring it.
- Skills: the procedures an agent follows.
- Autonomy and governance: what its level lets it do.