Claude's instructions for delegating work to subagents were rewritten

What

The Agent tool lets Claude hand part of a task to a subagent, a helper agent with its own working space. The instructions Claude receives about it changed:

• The Agent tool description was rewritten. It used to begin "When using the Agent tool, specify a subagent_type parameter to select which agent type to use. If omitted, the general-purpose agent is used." It now includes guidance such as "If research can be broken into independent questions, launch parallel forks in one message." This was seen in prompt captures across 14 model variants, and one other group of variants got a different edit.
• The same rewrite appears in the copy of the Agent tool description used with the CLAUDE.md reminder.
• The main system prompt's paragraph about the Agent tool was rewritten. It used to warn against using subagents excessively and against duplicating work. It now says to "Reach for it when research or multi-step implementation work would otherwise fill your context with raw output you won't need again."
• The Agent tool's description is now built with a modelMix value as well as the existing subagentBudgetNote.

Why

These are Claude's own instructions, so they shape how often and how it delegates. You may see Claude split research into parallel forks, or hand off bulky work, more readily than before.