creating-subagents

You isolate recurring tasks by creating specialized subagents with narrowly defined roles and scoped tool access. This delegates work to focused AI instances, reducing cross-task contamination and keeping domain-specific contexts separated.

The best way to orchestrate a multi-step AI workflow is to break complex processes into focused subagents, such as PM, Architect, and Implementer roles. This establishes clear handoffs and hooks between pipeline steps using isolated contexts.

You create a subagent for domain-specific data science tasks by defining a named configuration with required fields like name and description, then issuing an instruction to activate it with domain-specific isolation and a defined toolset.

Yes, you can spawn subagents that are reusable across multiple tasks or repositories. This improves consistency by applying the same isolated context and scoped tool access to recurring domain-specific work across different projects.

Scripts and references are optional components when building task-specific AI agents. You must enforce core fields like name and description, but you can add scripts, references, and assets to enhance the subagent's specialized capabilities.