work-superspeedlisted

# Superspeed Fan a task out across real parallel sessions, reconcile it warm, and leave a log that says how to do it better next time. ## Read this before using it First read `~/.claude/references/parallelization.md` § Choose and preserve the agent setup. Derive the setup from the actual chat before partitioning; reconcile stale saved choices before retries. The benchmarks and Claude-specific cache, permission and billing claims here are historical Claude evidence, not Astra measurements. Astra uses native Codex permissions and reports unknown telemetry. The advantage is real but narrow, and it was measured rather than assumed. On 2026-08-06, Claude Code 2.1.220, Apple M2 8-core, Max plan, four configurations and fourteen reps: | | one session, 4 subagents | one session, 4 teammates | **4 parallel `claude -p`** | |---|---:|---:|---:| | trivial task, 1 round | 52s | 59s | **22s** | | real task, 1 round | 64s | 59s | **31s** | | trivial, 3 rounds | 111s | 121s | **75s** | Separate sessions won every configuration and every individual rep, at identical correctness (24/24). **The mechanism is a fixed ~33s, not a multiplier.** The gap was 30s at one round and 36s at three, so it is a one-time orchestration toll, not something that compounds. That makes it 2.4x on a 22-second job and under 1% on an hour-long one. **So the honest rule: use this when a task splits into pieces that can run at once. Do not use it hoping to make one long serial task faster.** If the work does no