Better with super.engineering

Autopilot runs standalone on stock Claude Code, but it is built to sit on top of super.engineering and its sc CLI, and that is the recommended way to run it. Where sc is available it is the dispatch vehicle; Agent-tool dispatch is the degradation, not the default.

The orchestrator probes once per job — sc on PATH and sc agents list --output json succeeding — and if it answers, four things change:

• Roles become real agents. Each dispatch launches as a labeled sc agent (t-###-<role>) in its own pane through sc layout run, so you watch each role work instead of waiting on an opaque subagent. Follow-ups, retries, and review rounds go back to the same live agent with sc agent send.
• Cross-family review stops being aspirational. The adversarial panel resolves real providers from sc layout capabilities and deliberately seats a reviewer from a different family than the producer. That disjoint-findings result under Field results is what this buys.
• Fan-out survives a headless session. sc agents get resolves a callback address, so finished workers wake the dispatching session by queued message rather than by being waited on. Without one, a headless session keeps every dispatch in-turn.
• True isolation when paths overlap. Parallel code briefs touching the same files take sc worktree create instead of sharing one working tree — only ever on your explicit request.

None of it is required. Any sc failure, at probe time or mid-job, falls through to Agent-tool dispatch with the same brief and logs the degradation. Registry cli and mcp workers are unaffected either way.