Skip to content

Build your first agent

An agent is the Polo teammate that takes your tasks: its AI model, its instructions, its tools and its limits. You build one in the agent builder in three steps. Only a workspace administrator can create or edit agents.

VideoBuild an agent in three steps

Add at least one model under Settings → AI models. The builder cannot finish without a model to choose. See Models and spend.

You do not need to connect your systems first. The builder lists what a template needs that is not connected yet, and you can add it later.

  1. Open Agents.
  2. Choose Create agent.

On a new workspace, Home also offers Create an agent in the Ask Polo box.

Pick what the agent is for. Polo sets up its tools and limits, and you can change them before creating it.

Template What it is set up for
Monitoring agent Watches services and reports with evidence. Read-only.
Infrastructure agent Plans Terraform; applies only after approval.
Deployments agent Kubernetes and Helm releases and rollbacks.
Security agent Read-only scans of images, CI and inventory.
FinOps agent Read-only cost, inventory and waste reads; reports savings.
Pipeline agent Measures a slow pipeline, changes it in a draft PR and compares that PR’s CI with the baseline.
Start from scratch No tools or limits set. You choose everything.
  1. Enter a Name, such as “Monitoring agent”.
  2. Choose the Default model. This is the model the agent thinks with.
  3. Set the Model budget per task (USD). A task stops when it reaches this amount.
  4. Choose Next.

This step shows everything the agent will be allowed to do.

If the template needs a system that is not connected, it is listed here with a link to where you connect it. You can create the agent now, or connect the system in another tab and choose Check again.

Under Polo enforces:

  • Tick Read-only: observe and report, never propose a change to lock editing, checks and change tools. Only reading tools stay on.
  • Whether or not you tick it, every change outside the agent’s own workspace copy needs a person’s approval, or a standing approval an administrator set for that exact change.

Switch a tool off to take it away.

  • Source code, in the agent’s own workspace copy: Read source, Edit source, Run approved checks.
  • Team and services: Read workspace member directory, Read the service catalog.
  • Connected systems, tools and knowledge: one entry per action on each connection, such as reading a cloud account or planning Terraform. This list appears once you have connections.
  • Skills: choose Add a Terraform skill to let the agent plan and apply Terraform. See Add Terraform to an agent.

Open this section to change the defaults the template set:

  • Specialty and Instructions. The first sentence of the instructions is the role shown on the agent’s card. Polo reads all of it before every task.
  • Environments and Services. Leave them empty to allow any. When you list names, Polo refuses tasks for any other.
  • Forbidden actions, one per line.
  • Fallback model. When the default model cannot answer, the task continues on this model inside the same budget and says so. Without a fallback the task stops.
  • Specialist models, for helpers that ask for a particular class of model.

When the review is right, choose Create agent. Use Back to return to an earlier step.

The agent appears as a card on Agents. The card shows:

  • its role, model, budget per task and limits;
  • a short summary of what Polo enforces, such as Read-only or Changes need approval;
  • What it can and can’t do: the full lists under Can do, Needs approval for and Never does.

Choose Edit on the card to reopen the builder on that agent.

Above your agents is Polo, the supervisor. It takes every request that names no agent and hands each part to an agent that holds tools. An agent with no tools is listed as not on its team. See Agents.

Open Home, describe the outcome in the Ask Polo box, pick the agent or leave Polo decides, and choose Assign to Polo. See Give work to Polo.