Skip to main content

Set up an agent

An agent is a tool-using LLM you configure once and then reuse — either by chatting with it yourself (see Use the assistant) or by wiring it into an automation that invokes it automatically.

note

This is about the AI agents you configure in-app. If you're looking for the separate Kubernetes discovery agent (the one that runs in your cluster to find services), that's a different thing — covered elsewhere.

Create an agent​

  1. Go to Agents. The Agents tab lists what you've built; Tool Servers is where you connect MCP servers agents can draw tools from.
  2. Click + New Agent to start from scratch, or Browse Templates for a prebuilt one (for example, an Incident Manager template that also sets up the automation that triggers it).
  3. Fill in:
    • Agent Name and an optional Description
    • System Prompt — instructions defining the agent's role and behavior (up to 8,000 characters)
    • Max Iterations — how many tool-call rounds it can take before stopping (1–20, default 6)
    • Tool Allowlist — which of OpsRadar's built-in tools it can use
    • MCP Tool Access — which tools from your connected MCP servers it can use
  4. An agent needs a name, a system prompt, and at least one usable tool before you can save it.

The new agent form, filled in with a name, description, system prompt, and two tool allowlist entries checked

Governing what an agent can do​

Two independent layers control this, and both are enforced server-side, not just hidden in the UI:

Tool allowlist. Each built-in tool carries a required permission scope, is marked read or write, and has an approval mode — automatic or manual. Adding a write tool with manual approval means every call it makes waits for a human to sign off.

MCP connection scopes. For tools coming from external MCP servers, access is set per tool: Ask (approval every call), Ask once per conversation (covers roughly 30 minutes or 20 calls), Always allow, or Block. If an MCP server changes a tool's definition after you've granted access, OpsRadar automatically suspends that grant and falls back to asking again — it won't silently keep trusting a tool that changed underneath it.

Anything waiting on a human shows up in the Approvals inbox, grouped by conversation turn, so you can approve or reject a batch at once.

Scope​

Agents aren't scoped to individual services — access is organization-wide, governed by the tool allowlist and MCP connections above. There's currently no way to restrict an agent to only act on a specific service.

Triggering an agent​

Agents don't run on their own — something has to invoke them:

  • From the assistant — pick the agent in a chat and talk to it directly.
  • From an automation — wire it to a trigger (an incident opening, a deploy, a schedule) so it runs unattended. You can't hand-build this kind of automation directly; it's created automatically when you instantiate an agent template that includes one.