Subagent Forking Is On by Default. A Fork Inherits Your Mistakes at 10% of the Price.

Claude Code 2.1.232 shipped today with one line that changes the economics of delegation: "Subagent forking is now on by default: a subagent_type: "fork" subagent inherits the full conversation and prompt cache."

Price that out. A fresh subagent re-reads the files, rebuilds the context, and pays full input rate for every token of it. A fork reads that same prefix from cache at 0.1× the base input price — Anthropic's published multiplier. Briefing a delegate went from a line item to a rounding error. The same release also makes non-teammate agent spawns run in the background by default. Both defaults push the same way: delegate more, delegate sooner.

Here's the part that isn't in the changelog. A verifier subagent catches things because it starts cold — it re-reads the code instead of trusting the summary the parent wrote about it. A fork starts with the parent's full conversation, which means it starts with the parent's conclusions: the misread config, the wrong assumption from three hours back. It will agree with you faster and cheaper than a clean window ever could.

Forking is the right default for fan-out — several agents working independently over shared context. It's the wrong default for review. If your harness spawns verifiers, that's the one call that should still pay full price for cold eyes.

See how we build the guardrails →