Fable-agents splits Claude Code work across three agents

Separating implementation from judgment is a practical fix for the self-review problem that makes coding agents overrate their own output.

• –Fable handles orchestration while Sonnet performs implementation and Opus reviews the result.
• –The gatekeeper checks the diff against the original goal and runs real project verification, but never edits code.
• –Explicit model routing reduces the risk that every task silently consumes the most expensive model.
• –Global installation makes the setup reusable across projects, though it also increases the risk of prompt and configuration conflicts.
• –The architecture is more useful as a workflow pattern than as a turnkey autonomous engineering system: independent verification still depends on good tests and clear acceptance criteria.