Custom subagents

Today: ad-hoc subagents

Magic Coder is smart enough to formulate subagents on the fly when a task calls for it. If the agent decides a sub-task would benefit from running with a fresh context, a narrower scope, or a different posture (e.g., a focused reviewer over the diff, a search agent that scans the monorepo before answering), it dispatches one ad-hoc — without you having to declare anything in advance.

You don’t need to configure or register a subagent for this to happen. Just describe the task you want done; Magic Coder decides whether breaking it into sub-tasks is worth it.

Coming soon: declarative subagents

Support for declarative subagent definitions is on the way. Once it ships, you’ll be able to drop a manifest into your project (or user-level) directory and have it pick up any subagent you define — its system prompt, its tool set, and optionally its model and reasoning effort. The format will be Claude Code-compatible, so existing definitions port over without rewriting.

This page will be expanded with the manifest format, the dispatch protocol, and worked examples once the feature lands.