agent-design

Autonomous agents require isolated contexts with no embedded user prompts. Structure your agent with a minimal body (H1 and Process sections), load skills rather than embed knowledge, and craft precise delegation triggers in the description field so Claude autonomously executes tasks based on context.

Match tools to your agent's responsibilities using the principle of least privilege. Select from Read, Write, Bash, and other available tools based on what your agent actually needs to complete its task, ensuring efficient and secure execution.

Craft precise description fields that enable Claude to recognize when to delegate tasks to your agent based on conversational context. The description should clearly signal the agent's responsibility so delegation happens autonomously without user prompts.

Yes. Agent-design applies directly to sub-agent creation and review within box-factory-architecture, covering decisions on sub-agent versus command versus skill, agent body structure, skill loading, tool usage, and strict isolation requirements.

Agents must rely on loaded skills rather than embedded knowledge to maintain isolation and autonomous operation. Embedding knowledge breaks the isolated context principle; instead, structure your agent to load and invoke skills as needed.

Required sections are H1 and Process. Optional sections include Prerequisites, Skill Usage, Constraints, and Error Handling. This minimal structure enforces isolation and ensures your agent operates autonomously without user interaction.