fable-agent-orchestration
A Claude Code skill: an orchestration playbook for driving many coding agents through a large multi-PR engagement.
A collection of reusable workflow skills for coordinating multiple AI coding agents on large software projects, designed for use with Claude Code. A conductor manages scope and reviews while isolated coding agents each handle a small piece of work.
This repository is a collection of reusable workflow skills for coordinating multiple AI coding agents on large software projects. It is designed to work with Claude Code, an AI assistant. The core idea is that a human or lead agent acts as a conductor who manages scope, reviews work, and decides what gets merged. The individual coding agents each handle a small, isolated piece of the implementation work.
The main orchestration loop involves splitting work into independent slices, launching one agent per slice in an isolated git worktree, and giving each agent a detailed packet describing its role, scope, and proof requirements. Each agent must open a pull request rather than merge its own work. Two separate critic roles then review the submission, one checking the tests and one checking the code. Their claims are verified against real code and continuous integration results before anything is merged. The conductor can relaunch the next slice while reviews run.
The repository contains a single-file skill document, a public playbook, a machine-readable catalog, and a set of individual skill files. The skills cover topics like worktree isolation, preventing fake test results, recovering stalled work, controlling instruction drift, and running adversarial reviews. There are over twenty named skills in total, each addressing a specific failure mode in multi-agent engineering workflows.
The project is explicit about what it excludes. It does not contain private project formulas, identity material, secrets, session transcripts, or third-party source code. It credits three external MIT and Apache licensed projects as inspiration for specific skills, but states that no external code was vendored. All skills are rewritten syntheses for this repository's own orchestration model.
The repository is licensed under the GNU Affero General Public License v3.0 or later. Earlier versions published under Apache 2.0 remain available under their original terms. The README is thorough and clearly documents the boundary between public and private material.
Where it fits
- Coordinate multiple AI coding agents on a large software project by splitting work into isolated slices.
- Set up a review pipeline where critic agents verify test results and code quality before merging.
- Prevent AI agents from faking test results by requiring proof checked against real code and CI.
- Recover stalled agent work and control instruction drift across long multi-agent sessions.