The prompt cache stopped being a pricing feature and became a scheduling problem

Claude Code 2.1.229 shipped a config flag most people will scroll past: CLAUDE_CODE_WORKFLOW_PREFIX_STAGGER_MS.

What it does is delay the launch of sibling agents that share a prompt prefix, so the second through Nth agents read the warm cache instead of each writing their own copy of the identical preamble (changelog). Set it to 0 and you get the old simultaneous behavior. Previously, if you fanned out ten subagents with the same 40K-token system context, all ten paid cache-write pricing on that prefix, because they all hit the API before any of them had populated the cache.

Three independent stacks landed the same insight within days of each other. Block's goose v1.46.0 shipped cache-safe request assembly using append-only turn context with declared cache semantics (release notes, PR #11022), specifically to keep hit rates up on long sessions. Headroom v0.35.0 attacks it from the other end with token compression plus explicit prompt-cache TTL pricing in the CLI (release, PR #2802). And Terminal-Bench 3.0 now ranks agents on cost and token usage alongside resolution rate (tbench.ai).

The through-line: with fan-out orchestration normal, the dominant cost driver is no longer which model you picked. It's how many agents redundantly pay for identical context. Simultaneity is the anti-pattern. Launch parallel siblings with a deliberate offset.

That's a one-line config change and it's the highest-ROI thing in this newsletter. If you're running Claude Code workflows with fan-out, set the stagger, then check your cache-read versus cache-write ratio before and after. Related fix in the same release worth knowing: dynamic workflows were sizing concurrency off the host core count inside CPU-limited containers, so a 2-core cgroup on a 16-core host was spawning 16-way parallelism. If your containerized fan-outs have been thrashing, that's why.

The benchmark shift makes this permanent. On Terminal-Bench 2.1, Qwen3.8-Max (86.6), GPT-5.6 Sol (88.8), Opus 4.8 (84.6), and Fable 5 (84.6) sat inside a four-point band. When pass rates converge to noise, the argument moves to tokens-per-resolved-task. Start recording token spend in your own evals now, before someone asks you to justify the bill.