flow

Decision 0031 · Accepted

Fan out
by default

Flow's skills are mostly written as sequences.

Status
Accepted
Built from:
da88c48

Status: Accepted (2026-10-28). Amends 0025 and 0029 (one review worker per spec becomes one worker per axis plus a verifier). Spec: 019. Amended by 0042 (parallel workers and one review worker per axis do not bind build, ship, or the review they run). Amended by 0043 (item 1: workers only where they pay). Amended by 0044 (independent work runs in parallel by default again; the agent chooses how). Amended by 0053 (parallel when it pays, not by default). Amended by 0060 (item 5: areas also slice specs without tasks; build sends each slice to a worker).

Context

Flow's skills are mostly written as sequences. write-agents and review use separate workers only when a task "justifies" them, ship runs its review with one worker, and explore finishes all reading before its first question. Only build runs work in parallel. Hosts now offer cheap parallel workers, and the user wants maximum parallel execution for speed.

Decision

  1. Parallel is the default. A skill launches its independent steps as parallel workers in one batch, up to the host's limit. A step stays sequential only when it needs another step's output, asks the user, or writes a shared file.
  2. One home for the rule: write-agents Workers. Skills name their parallel groups and link to the rule rather than restating it.
  3. Workers are bounded. Each is read-only or writes files no other worker touches. The main thread verifies and synthesizes, and asks the user every question.
  4. The spec review splits by axis. It runs one worker per review axis the change can affect, then one verifier. Together they are the one independent review per spec.
  5. Tasks declare an area (package or directory), so build can schedule non-overlapping tasks together.
  6. Fallback. Without workers, the same steps run in order, and the skill says so.

Alternatives

  • Keep workers opt-in per skill. Rejected: skills stay sequential by default, and speed depends on the agent noticing what can run in parallel.
  • A subagent per step everywhere. Rejected: single-pass skills (handoff, tldr, interview) lose context and gain nothing.
  • Infer overlap from file paths in tasks. Rejected: paths go stale; a coarse area is stable enough to schedule on.

Consequences

  • Faster discovery and review. Token use rises with the number of workers.
  • Skills gain a short parallel-groups line, and their copies of the worker rule are removed.
  • Reviews on hosts without workers are labelled single-agent, as before.

Jump to, or try a question