BetaYou are reading docs for the beta channel (
2.13.0-beta.20). Features may change before the next stable release.Switch to stable →Understand AgentKit
Learn the product model behind kits, runtime adapters, projects, artifacts, and recovery.
AgentKit brings structured AI workflows into the coding assistant you already use. These pages explain what AgentKit installs, how a runtime uses it, and where your responsibility begins and ends.
Start with the mental model
- AgentKit architecture — follow a kit from a verified release to your coding assistant and then to a reviewed result.
- Kits, Skills, Agents, and Hooks — understand the job of each building block and where to start.
- Runtime adapters — see why the same kit behaves differently in Claude Code, Codex, and Cursor.
- Projects, artifacts, and checkpoints — understand scope, file ownership, preserved work, snapshots, and review gates.
Three ideas to keep in mind
- A kit is installed; a workflow is invoked. Installing makes supported components available. It does not start a task by itself.
- The runtime remains in charge of execution. AgentKit adapts content to the runtime, but the runtime controls model access, tools, permissions, and event support.
- You remain in charge of the result. Installation metadata and recovery snapshots reduce risk. They do not replace reviewing plans, diffs, generated files, or external actions.
When you are ready to use these ideas, continue to Installing kits.