liza-mas/liza

What it solves

Liza addresses the problem that AI coding agents often fake progress instead of solving problems, and that reviewing everything they do consumes scarce human attention. It is a system for software delivery where agents decompose a goal, produce specifications and code, review each other's work, and merge approved changes through code-enforced gates. The key metric is "cost to production-ready": tokens plus human time to reach a mergeable change, including correction, tests, documentation, and review.

How it works

Liza uses four layers to keep execution reliable:

• A behavioral contract makes agents reason before acting, admit uncertainty, verify claims, and push back when warranted. It makes saying "I'm stuck" safe instead of motivating fake progress. An Approval Request mechanism forces agents to write down their reasoning before acting.
• Adversarial doer/reviewer pairs apply a binding review to every artifact, from epics to code. A reviewer must Accept, Counter, Refute, or Escalate findings; a doer can contest a finding by naming the concrete harm its fix would cause. Repeating the same assertion is not a resolution.
• Mechanical boundaries are deterministic Go code that enforce state transitions, role boundaries, validation gates, and merge authority. These rules do not depend on the model choosing to follow them.
• An engineered pipeline decomposes goals into cohesive, verifiable tasks whose outputs compose. The shipped pipeline covers specification, architecture, coding, and integration, with doer/reviewer pairs throughout.

The system runs provider coding CLIs (like Claude Code and Codex), so it uses existing subscriptions. It adds isolated Git worktrees, an auditable YAML blackboard, pre-claimed tasks, a live TUI, crash recovery, context handoff, and a circuit breaker for repeated failures.

Liza has three modes. Pairing is direct human-agent collaboration under the contract. Adversarial pairing adds a separate reviewer session with a dedicated worktree and binding verdicts. Full MAS (multi-agent system) uses deterministic supervisors, a task graph, worktrees, and thirteen roles for goals needing decomposition and unattended sprints.

Human attention goes to direction, domain knowledge, and constraints. Within a sprint, agents execute and review autonomously. Between sprints, humans resolve domain questions and steer the next increment.

Who it's for

Liza is for engineers and teams who want to delegate software delivery to AI agents without giving up governance. It suits people with existing Claude Code or Codex subscriptions who have meaningful validation coverage and want unattended sprints, or who simply want a stronger behavioral contract in one-on-one pairing. It is complex and requires reading the setup guides before first runs.

Highlights

• Behavioral contract that makes agents reason before acting and admit uncertainty, reducing fake progress
• Adversarial doer/reviewer pairs for every artifact, with binding verdicts that humans can override between sprints
• Deterministic Go code enforces state transitions, role boundaries, and merge authority — rules do not depend on model cooperation
• A pre-engineered graph for software delivery that is instantiated per goal and adaptable during execution, so users do not have to design agent workflows each time
• Three modes: Pairing, Adversarial Pairing, and Full MAS, so the layers can be adopted to match the work
• Uses existing provider CLIs and subscriptions (Claude Code, Codex), with configurable provider selection
• Supports crash recovery, context handoff, a circuit breaker for repeated failures, and a live TUI