DISCOVER THE FUTURE OF AI AGENTSarrow_forward

Liza

calendar_todayAdded Feb 25, 2026
categoryAgent & Tooling
codeOpen Source
Multi-Agent SystemGoAI AgentsAgent FrameworkCLIAgent & ToolingDeveloper Tools & CodingAutomation, Workflow & RPA

A peer-supervised multi-agent coding system built on behavioral contracts. Enforces Tier 0 invariants to constrain AI coding agent behavior, supports Planner-Coder-Reviewer role separation with external validation, transforming AI coding agents from "eager assistants" into "reliable senior engineering peers".

Project Positioning#

Liza is a multi-agent coding system built on behavioral contracts, with the core goal of transforming AI coding agents from "eager assistants" into "reliable senior engineering peers". The project achieves this through four pillars: Behavioral Contracts, Specification System, Blackboard Coordination, and External Validation.

Core Capabilities#

Behavioral Constraint System#

  • Tier 0 Invariants: Prohibits unapproved state changes, hallucination, test tampering, unverified success claims
  • Tier 1-3 Degradation: Behavior constraints can be degraded by level under pressure

Coordination Mechanisms#

  • Blackboard Coordination: Shared state.yaml tracks goals, tasks, assignments, history
  • Leases Mechanism: Task lease system with expiration and recovery
  • Hypothesis Exhaustion: Auto-triggers task restructuring when two coders fail same task

Quality Assurance#

  • External Validation: Coders cannot self-certify, must be approved by Code Reviewer
  • Commit SHA Verification: Coder submits SHA, Reviewer verifies before review
  • Approval-gated Merge: Merge only after Code Reviewer approval
  • 43+ Validation Rules: Strict state machine transitions

Agent Roles#

RoleResponsibilitiesProhibited Actions
PlannerDecompose goals into tasks, restructure failed tasksImplement code
CoderClaim tasks, iterate implementation, TDD developmentSelf-approve or merge
Code ReviewerReview work, approve/reject, mergeImplement code

Task Lifecycle#

DRAFT → READY → IMPLEMENTING → REVIEWING → APPROVED → MERGED
             │              │
             │              └─> REJECTED ──┐
             ├──> BLOCKED ──> READY (rescoped)
             │                ├──> SUPERSEDED
             │                └──> ABANDONED
             └──> INTEGRATION_FAILED ──┘

Deployment Modes#

  1. Pairing Mode — Human-AI pairing with human approval gates
  2. Multi-Agent Mode — Multi-agent collaboration with peer supervision
  3. Subagent Mode — Delegate subagents for focused tasks

Quick Start#

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

# Verify Installation
liza version

# Initialize Project
liza setup
liza init "Project goal" --spec specs/vision.md

LLM Compatibility#

  • ✅ Claude Opus 4.5 (Reference Provider)
  • ✅ GPT-5.2-Codex
  • ⚠️ Mistral Devstral-2 (Requires explicit activation and supervision)
  • ❌ Gemini 2.5 Flash (Architecture limitation)

System Requirements#

  • Claude Code or Codex CLI
  • Git 2.38+ (Full worktree support)
  • Go 1.25.5+ (Source build only)

Design Philosophy#

"The contract defines what's forbidden; the shape that remains is where judgment lives."

Related Projects

View All arrow_forward

STAY UPDATED

Get the latest AI tools and trends delivered straight to your inbox. No spam, just intelligence.

rocket_launch