01

Parallelize independent evidence, not shared uncertainty

Supported agent systems can delegate bounded work, but simultaneous edits can conflict and additional workers consume resources. The ownership matrix below is our project-management method. It does not imply automatic communication between Claude Code and Codex.

02

Before you start

Have a stable brief and a controller who can integrate results. If the work is one tiny label change, a single worker is usually easier to review. Parallelism is useful only when there are independent tasks worth the coordination cost.

03

Example: a larger form cleanup with three owners

Worker A owns contact.html and its markup tests. Worker B reviews keyboard behavior read-only and returns findings. The controller owns integration and the final browser pass. Nobody is allowed to rewrite the shared stylesheet until ownership is assigned.

  • Dependency gate: B reviews the exact patch A marks ready, not a moving half-finished state.
  • Budget: one implementation pass and one review pass before a checkpoint; ask before expanding scope.
  • Deliverable: each worker returns files, checks, findings, and limitations.
  • Conflict rule: if two tasks need the same file, stop and reassign it rather than racing to save.
04

Work through the task

  1. Draw a task list with inputs, outputs, allowed files, and dependencies. Remove tasks that cannot run independently.
  2. Assign one write owner per file or isolated checkout; give reviewers read-only boundaries.
  3. Set a checkpoint and a stop rule for repeated failure, missing permissions, or expanding requirements.
  4. Integrate one reviewed result at a time and rerun shared checks on the combined state. A worker’s passing test applies only to the state it tested.
05

What a useful result looks like

Parallel work should produce independently inspectable artifacts and a final integrated result. The controller remains responsible for gaps between tasks, such as a markup change that invalidates a browser selector used by another worker.

06

When to stop or change direction

Stop a worker when its input contract changes, when it touches another owner’s files, or when repeated retries add cost without new evidence. Do not keep idle agents running just to make the workflow look parallel.

07

Verify before continuing

  • Does every editable file have exactly one owner?
  • Are dependency gates tied to a real revision?
  • Did the combined result receive verification after integration?

QUESTIONS THIS ANSWERS

Questions this answers

  • How do I prevent parallel agents from editing the same file?
  • When is parallel AI work worth the coordination cost?

Found something wrong? Report an error or read the corrections policy.

2 SOURCESEvidence ledger

Sources

  1. 01
    Subagents ↗

    OpenAI · accessed 4 Sept 2026

  2. 02
    Create custom subagents ↗

    Anthropic · accessed 4 Sept 2026