Core Concepts

AgentTeams runs on four concepts. Once you know them, the whole flow makes sense.

Plan

A unit that defines the intent and scope of work. The agent drafts it; once a human reviews and approves, execution begins. See Plans for details.

Agent

The AI agent that performs the work. You can run several, organized by role (e.g. claude-backend, claude-frontend, claude-review).

An agent is registered automatically when you run agentteams init in your project root — the CLI detects the environment (Claude Code · Codex · OpenCode, etc.). See Agent Configuration for registration and management.

Runner

A local executor that picks up requests and runs the configured agent. It registers as an OS service and starts automatically, and you can run one on each machine.

Runner is part of the automated execution flow, but it is not required for first-time setup. The 5-Minute Quickstart covers the minimum path: initialize the CLI and connect a local AI agent. Install a runner later when you want AgentTeams requests to be picked up and executed automatically. See Runner Installation and Runner Requests for details.

Convention

The project rules and guides an agent must follow when writing code or doing work — code style, prevention rules, and more. See Conventions for details.

How it fits together

A human approves a Plan, the Runner runs an Agent, and the agent works according to Conventions. On success a Completion Report is produced; on failure a Post-mortem, whose prevention measures feed back into Conventions.

When you’re ready, try it yourself in the 5-Minute Quickstart.