agent-authoring

Write a description field that leads with the use case and names concrete trigger phrases, using the "proactively" convention when you want automatic delegation. Claude decides whether to delegate by reading the description, so vague descriptions mean the agent is never used.

Specify a least-privilege tools allowlist such as Read, Grep, Glob for read-only analyzers, adding Edit or Write only when the mandate requires it. Pair it with the weakest permissionMode that still completes the job, reserving bypassPermissions for trusted narrow automation.

Choose an agent when a self-contained side task produces noisy output and can return a tidy summary in an isolated context. Choose a skill for reusable in-conversation expertise, a command for one-shot prompts, and a hook for deterministic event-triggered actions.

The description is almost certainly too generic, since delegation decisions are driven entirely by that field. Sharpen it with specific situations and trigger phrases before changing tools, permissions, or the body.

No, a subagent starts fresh with a separate context window and should not be assumed to absorb every CLAUDE.md rule. Anything it must know must be restated in its system-prompt body or passed in its task input.

Match the model to the reasoning load: haiku for fast mechanical or high-volume work, sonnet as the default for most agents, opus for complex multi-step reasoning, or inherit to follow the parent session's model.