modu-ai/moai-adk

What it solves

MoAI-ADK is a verification-driven agent orchestration harness designed to make AI-generated code trustworthy. It solves the problem of "hallucinated" completion claims—where an agent claims tests passed without actually running them—by banning unverified claims at the system level and binding every success claim to actual command output as evidence.

How it works

It acts as a structural wrapper around Claude Code, managing the orchestration of models, reasoning depth, and verification. It implements a structured SPEC lifecycle (plan $\rightarrow$ run $>$ sync) and uses a "Kanban Mode" to split work across multiple terminals (lead, plan, run, and sync sessions). This prevents context window saturation by ensuring each session only carries the context relevant to its specific phase. It also utilizes isolated worktrees and branch guards to allow parallel-safe execution without agents stepping on each other's code.

Who it’s for

It is built for developers and engineering teams using agentic coding tools who need strict quality control, cost efficiency (tokenomics), and the ability to manage long-horizon tasks that survive session clears.

Highlights

• Evidence-Based Verification: Forbids unverified completion claims; every claim must be bound to a command run and its output.
• Kanban & Factory Modes: Orchestrates work across multiple specialized sessions to optimize context usage and allow parallel processing.
• Parallel-Safe Isolation: Uses dedicated worktrees and branch-state guards to prevent accidental code overwrites during parallel runs.
• Tokenomics Optimization: Declaratively assigns models and reasoning depth based on the work phase to reduce token spend (e.g., using cheaper models for implementation and high-reasoning models for planning).
• Autonomous Boundaries: Implements hard limits on turn counts, wall-clock time, and stagnation guards to prevent infinite loops during autonomous execution.
• Self-Improving Loop: Captures failure patterns and routing decisions into a decision memory to improve future runs.