This repository contains the whitepaper Trained Monkeys: A Methodology for AI-Orchestrated Software Development.
This paper presents a structured methodology for AI-orchestrated software development using a three-stage pipeline: planner, decomposer, and supervisor. The approach separates high-cost reasoning from low-cost execution by reserving the expensive model for planning, decomposition, and verification while delegating implementation steps to cheaper worker models. It defines machine-consumable artifacts, strict role separation, dependency-aware parallel execution, and a zero-trust verification protocol in which worker outputs are treated as hints and validated against ground truth. The methodology further specifies bounded retries, revision loops, and escalation paths to improve operational reliability. The contribution is descriptive and practice-oriented, offering an auditable framework for adopting multi-agent development workflows under real-world cost and context-window constraints.
The paper describes a way to run software development with AI that stays affordable, auditable, and under human control, instead of one agent doing everything in a single long conversation.
Pipeline (each stage driven by a high-capability model):
- Planner — sets the goal and phases
- Decomposer — breaks each phase into small, self-contained steps
- Supervisor — hands steps to inexpensive worker agents running in parallel, then independently verifies every result before accepting it
All coordination happens through versioned files in the project repository, so plans, progress, and outcomes are recorded rather than lost in a chat transcript.
Benefits:
- Lower cost
- Verifiable results
- Scale beyond a single agent's limits
- Fault tolerance
The three skill definitions the paper describes, in the claude-skills repository (included here as the claude-skills git submodule):