I Gave Codex 3 Jobs at Once—Then Checked Every Merge

Give each worker a bounded contract

Search, summary, and snapshot restore have different responsibilities, so they can be assigned independently. They still depend on the same task representation. The experiment starts the worktrees from a common commit and makes that starting point visible. Describe expected behavior and the fields each task may rely on before workers begin changing the project.

Review progress as an incomplete whole

The published common suite begins at 3 of 21 passing checks, then reaches 8, 11, and 21 as the reviewed changes enter the main checkout. An intermediate result can be correctly incomplete when another worker’s responsibility has not yet been integrated. Inspect the failed cases and the diff instead of treating every partial score as a new failure by one agent.

Look beyond successful parsing

The summary cases include a zero-minute task, while snapshot restore has to do more than accept parseable input. The episode checks bad input as well as the normal path. Those details make the shared contract useful: preserving a valid zero and restoring the intended state are behaviors that broad instructions such as improve reporting can leave ambiguous.

Keep the speed claim separate from the experiment

This was an actual Codex CLI run with manually created branch-based worktrees. The app-managed worktree and Handoff documentation describes a different setup path. No matched sequential comparison was run, so the episode does not establish a threefold speed gain. The result covers three functions in an owned synthetic repository, with no deployment or whole-app security audit.

Jump to a chapter