deep-agents-core

You can bootstrap a multi-agent application by using a centralized harness that configures built-in middleware like TodoListMiddleware and SubAgentMiddleware. This provides a reusable blueprint to orchestrate tasks, manage memory, and delegate to subagents consistently across environments.

On-demand skill loading dynamically imports agent capabilities from a YAML frontmatter SKILL.md file. This progressive disclosure pattern ensures only necessary skills are loaded at runtime, reducing memory overhead and keeping the agent lightweight.

You coordinate tasks and manage memory by configuring the agent harness with built-in middleware. TodoListMiddleware tracks task progress, MemoryMiddleware handles persistence, and FilesystemMiddleware manages data, allowing complex agents to operate consistently.

Yes, the harness supports scalable deployments across development and production environments. It allows customizable prompts, subagent delegation, and persistence backends, ensuring consistent task coordination and data management for team workflows.

Yes, a SKILL.md file with YAML frontmatter containing a name and description is required. This file enables the on-demand skill loading mechanism, supports optional scripts and references, and ensures standardized skill discovery.

The best way to structure complex agents is using a centralized harness with SubAgentMiddleware. This approach enforces standardized skill loading via SKILL.md and coordinates tasks through built-in middleware, ensuring scalable and maintainable agent architectures.