Overview
Managed Agents is Cluster's runtime for autonomous AI agents — define an agent once, then run it on demand, on a schedule, or from a GitHub event.
Managed Agents
Managed Agents is Cluster's runtime for autonomous AI agents. You define an agent once — a model, a system prompt, a set of tools — and then run it however you need: interactively, on a cron schedule, in response to a GitHub event, or as a one-off job in a fresh VM.
Agents run on the managed-agents service at https://agents.clusterbase.dev.
You can drive everything two ways:
- The Console — a visual agent builder, agent editor, session transcripts, schedules, and triggers. The fastest way to get started.
- The HTTP API — the same capabilities over a REST API, for building agents into your own products and pipelines. See the API reference.
The model
Three objects make up the system. Understanding how they relate is most of the mental model:
| Object | What it is |
|---|---|
| Agent | A reusable, versioned configuration: name, model, system prompt, tools, MCP servers, reasoning effort, and an optional default VM. |
| Session | A single runtime instance of an agent. It holds the conversation, pins the agent version it was created with, and tracks token usage. |
| Event | One message on a session's timeline — your input (user.message), the agent's output, tool calls, tool results, and status changes. |
You create an agent, start a session on it, append a user.message
event, and the agent runs — emitting thinking, tool calls, and messages back
onto the session as events you can stream live. See Concepts
for the full domain model.
What you can do
- Build and configure agents — pick a model, write a system prompt, attach tools and MCP servers, and (optionally) give the agent a VM for shell and file access.
- Run agents interactively — start a session, send messages, and stream the agent's reasoning, tool calls, and replies over SSE.
- Schedule runs — fire an agent on a cron schedule in any timezone (e.g. publish a daily report every weekday at 9am).
- Trigger on GitHub events — run an agent when a PR
opens, or mention
@cluster-build reviewin a PR for an instant code review. - Launch on-demand runs — provision an ephemeral VM with your repos cloned, run the agent, and tear it down automatically.
- Provision environments — long-lived Firecracker VMs an agent's sessions execute inside.
- Store MCP credentials — connect agents to remote MCP servers, with secrets held in encrypted vaults.
Authentication
Every API request is authenticated with a Bearer JWT from
Sign in with Cluster (the same accounts.cluster.app identity
used across the platform). Agents, sessions, schedules, triggers, environments,
and vaults are owner-scoped: you only ever see and act on the resources
created under your own identity. See Authentication
for details.
Next steps
Start with the quickstart to build your first agent in a couple of minutes, then read Concepts for the full picture.