Agents

An agent is a named Polo teammate: its AI model, its instructions, the tools it may use, the changes that need your approval and what it must never do. The Agents page lists them as cards. Administrators create and edit agents; everyone can read them.

Above your agents the page shows Polo, the supervisor. Polo has no tools of its own. It takes every request that names no agent, gives each part to an agent that holds tools, and reviews the result. Each part runs with that agent’s own tools and rules. An agent with no tools is not on Polo’s team.

Create an agent

You need at least one model on Settings → AI models first.

1. Open Agents and press Create agent.
2. Choose a template. Polo sets up its tools and limits; you can change them before creating it.
3. Name and model. Enter a Name, choose the Default model and set the Model budget per task (USD).
4. Review and create. Check what it never does and the tools it may use, then press Create agent.

If the template needs something that is not connected yet, the dialog lists it under Not connected yet. Create the agent now and connect it later, or connect it in another tab and press Check again.

Templates

Template	What it is 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 pull request and compares that pull request’s CI with the baseline.
Start from scratch	No tools or limits set. You choose everything.

A template switches on only the tools that match something already connected in your workspace.

Tools are switches

Under Tools it may use, every tool is a switch. Switch one off to take it away. Tools are grouped: source code in the agent’s own workspace copy (read, edit, run approved checks), team and services (the member directory, the service catalog), the checks it may run, and connected systems, tools and knowledge.

• A tool that only reads runs without approval.
• A tool that makes a change is labelled after approval. The change waits for a person, or for a standing approval of that exact change. See Approvals and guardrails.
• Each tool that makes changes needs the read tool that verifies it. The dialog asks you to tick that read tool too.
• Read-only: observe and report, never propose a change locks editing, checks and change tools. Only its reading tools stay on.

Tools from systems you have not connected do not appear. Add them on Connections or, for MCP servers, on the Tools page, then edit the agent.

Skills

A skill is a packaged way of working that you add to an agent.

• Terraform skill. Press Add a Terraform skill and give it a name, the GitHub repository, folder and branch, the environment it changes, and the cloud account whose credentials Terraform runs with. For a private repository, store a GitHub access token in the vault and choose it. Tick Let it apply, after a person approves the plan to allow applies, and Let approved plans remove or replace resources only if you want that. State stays in the backend your repository declares. See the guide Add a Terraform skill.
• Knowledge skills. A document of type Skill on the Knowledge page can be granted to an agent at an exact version. See Knowledge.

Instructions, scope and limits

Open Customize instructions, limits and models for the rest.

Setting	What it does
Specialty and Instructions	What the agent is for and how it should work. Polo reads all of it before every task; the first sentence is the role shown on the card.
Environments, Services	Where it may work. Leave empty to allow any. Polo refuses tasks for any other environment or service.
Forbidden actions	Things the agent is told never to do, one per line.
Fallback model	The model a task continues on when the default cannot answer. The task says when it switched. Without one, the task stops.
Specialist models	A different model for helpers that ask for a named model class.

Caution

Forbidden actions are told to the agent, not enforced. A tool you grant can still be called. To block an action, keep the agent read-only, narrow its environments and services, switch the tool off, or rely on approval.

Budgets

Every task has an AI cost limit taken from its agent’s Model budget per task (USD). A task, a schedule, a monitor or an intake source can set a lower or different budget of its own. The agent’s card also shows the limits Polo enforces on each task: a token limit, a time limit that includes tool time, and how many agents one task may use. When a limit is reached the task stops and says so. How cost is counted is described in AI models and spend.

Read an agent’s card

Each card shows the agent’s role, its model (and fallback), its tools, a Needs approval for list, what it is told to avoid, and What it can and can’t do: the guardrails Polo enforces, such as Changes need approval, Read-only tools, or the environments and services it is limited to.

Edit an agent

1. Press Edit on the agent’s card.
2. Change anything, then save.

If someone else changed the agent while you had it open, Polo asks you to press Load the latest version and review it before saving. Changing an agent’s tools does not widen a monitor or schedule that already pinned its tools; update those where they are set up.

Limits

• Up to 12 environments and 24 services per agent.
• Up to 12 forbidden actions of up to 160 characters each.
• Instructions under 4,000 characters; a custom specialty up to 60 characters.
• The fallback model must differ from the default model.

Next: Create your first agent.