Unclear Claude Code's standard Agent tool text is recorded as unchanged in this release, so it is not clear which setups receive the rewritten guidance.
The Agent tool lets Claude hand part of a task to a separate helper agent and get a summary back. The text that tells Claude when to use it has been rewritten. The old text encouraged Claude to start agents when a task matched an available agent type, when work could run in parallel, or when answering would mean reading several files.
The new text says:
- a fresh agent knows only what is in its prompt, and Claude sees only its summary, so detail is lost both ways
- Claude cannot watch an agent work, only wait or cancel
- an agent's mistakes sound as confident as its findings, and an agent given a guess tends to confirm it
- several agents at once spend tokens (the units of usage you are billed for) in a burst you did not ask for
It asks Claude to weigh those costs against the accuracy agents bring. You pay for agents Claude did not need, and you pay again for work Claude redoes because it skipped one.
These are Claude's own instructions, so they change how Claude behaves. Judging from the wording, Claude may start agents less freely and treat what they report back with more care.
Claude Code's standard Agent tool text is recorded as unchanged in this release, so it is not clear which setups receive the rewritten…