liza-mas/liza

What it solves

Liza addresses the problem of "faking progress" and unreliable output in AI-driven software development. It aims to reduce the "cost to production-ready" by minimizing the need for constant human correction and vigilance, moving away from sycophantic AI assistants toward accountable, senior-peer-like agents that can admit uncertainty and push back when necessary.

How it works

Liza implements a multi-agent system (MAS) for software delivery based on a pre-engineered task graph. It uses four layers of protection to ensure quality:

1. Behavioral Contract: A set of obligations that force agents to reason before acting, verify claims, and admit when they are stuck.
2. Adversarial Pairs: Every artifact (from epics to code) is produced by a "doer" and verified by a "reviewer" to prevent the author from being the sole judge of their work.
3. Mechanical Boundaries: Deterministic Go-based enforcement of state transitions, role boundaries, and merge authority.
4. Engineered Pipeline: A structured decomposition of goals into specifications, architecture, and code plans that ensure outputs compose into a working system.

Who it’s for

Liza is designed for developers and engineering teams who want to automate software delivery pipelines while retaining governance and steering authority, specifically those looking for unattended execution of sprints without sacrificing production-grade craftsmanship.

Highlights

• Multi-Mode Operation: Supports Pairing (human-agent), Adversarial Pairing (doer/reviewer sessions), and Full MAS (autonomous task graph).
• Provider Agnostic: Works with existing subscriptions via coding CLIs, supporting models like Claude Opus and GPT-5.
• Binding Review Protocol: A structured PR-like dynamic where reviewers can Accept, Counter, Refute, or Escalate findings.
• Operational Tooling: Includes isolated Git worktrees, a YAML blackboard for auditing, a live TUI, and crash recovery.