Native implementation round

The idea

Implementation runs as one sandboxed Claude Code or Codex session, briefed by a compiled file. Its tools go through the gateway hook. The supervisor gets back only the report, the diff and the checks.

The problem
Relaying each tool call was slow: 164 calls on one replay, where the native run took a sixth of the time.
Why it matters
The delegate works many turns in one process. The employee keeps the context, the side effects and the verification.
Where it stands
In production as of

A step-by-step replay of this idea is not published yet. The library adds one when a recorded run shows it.

  • Employee supervises its agents

    The AI employee plans, briefs, reviews, integrates and talks to people. It hands bounded plan slices to cheaper models or agent CLIs.

  • Reusing the checks a delegate already ran

    After a delegated coding round, the platform reads which checks the delegate ran from its own record of that round. It reuses those results only when the code is unchanged, and runs the checks again otherwise.

  • Check-before-finishing Stop hook

    In a native Claude Code round, a Stop hook that uses no model holds the round once if code changed after the last passing check. A piped or masked status does not count as a pass.

  • Runner sandbox and egress proxy

    On the person's machine, a slice's CLI runs under sandbox-exec or bubblewrap, which deny by default. It writes only its task folder, and its only network is a proxy that reaches the vendor.

  • Host rounds on a person's own login

    A person's own CLI logins on the platform host can run rounds, but only for that person's own work. The host owner must first turn this on and accept the vendor policy.

Turn the numbers into shipped work.

Crewshal runs these choices for you: a persistent AI worker with memory and rules, on your Claude and Codex subscriptions.