Liza: Hardened Multi-Agent Coding System

Liza orchestrates LLM coding agents with behavioral contracts and adversarial review pairs to deliver production-quality code on the first pass.

Liza is a CLI tool — a hardened Multi-Agent Coding System that wraps LLM coding agents — Claude Code, Codex, OpenCode, Kimi, Mistral, and Gemini — with behavioral contracts, adversarial doer/reviewer pairs, and deterministic Go supervisors. Instead of hoping agents do the right thing, Liza enforces it mechanically, so you get production-quality code on the first pass, with full auditability to prove it.

​

Explore Liza

Quickstart

Install Liza, run global setup, initialize a project, and launch your first session in under 10 minutes.

Operating Modes

Understand the three modes — Pairing, Adversarial Pairing, and Multi-Agent — and when to use each.

Multi-Agent Mode

Run a fully autonomous spec-to-code pipeline with 13 roles across four pipeline phases.

CLI Reference

Every liza command, flag, and environment variable in one place.

​

Get Up and Running

1. Install Liza

   Run the install script on macOS or Linux (WSL2 on Windows):

   curl -fsSL https://raw.githubusercontent.com/liza-mas/liza/main/install.sh | bash
   

   Verify the installation with liza version.

2. Run global setup

   Set up contracts, skills, and provider integrations once per machine. Pass the flag for each provider you use:

   liza setup --claude --codex
   

   Then customize ~/.liza/AGENT_TOOLS.md to match the tools actually installed in your environment.

3. Initialize a project and start coding

   Navigate to your repository and initialize Liza for that project:

   liza init --claude
   

   Open the TUI to monitor agents and manage the pipeline, or start a Pairing session directly with your coding agent CLI:

   liza tui
   

​

Choose Your Mode

Pairing

Human-supervised collaboration. You stay in the loop — the agent codes under the behavioral contract while you steer, review, and approve.

Adversarial Pairing

A doer agent writes; a reviewer agent challenges through a shared Markdown blackboard. Higher confidence without the full autonomous pipeline.

Multi-Agent

Fully autonomous spec-to-code. Thirteen roles across four phases execute in parallel with adversarial review pairs on every task and human checkpoints between sprints.