Summary

Liza provides a hardened, auditable multi-agent coding system designed for production readiness. It targets engineers seeking reliable, maintainable AI-generated code, enhancing time-to-quality through disciplined automation.

How It Works

Liza utilizes a hybrid architecture: LLM coding agent CLIs are wrapped by deterministic Go supervisors. Its core is a "behavioral contract" addressing 55+ LLM failure modes, enforced intrinsically. An adversarial review process uses dual "doer/reviewer" agent pairs for all tasks, simulating PR workflows. Supervisors mechanically enforce state transitions and role boundaries, ensuring reliability and auditability beyond prompt-based systems.

Quick Start & Requirements

Installation involves downloading the liza binary, running liza setup, customizing AGENT_TOOLS.md (defining LLM tools), and initializing projects with liza init. Requirements include the liza binary, Git, Go, and compatible LLM provider CLIs (e.g., Claude, ChatGPT).

Highlighted Details

• Classified as L4 (Collaborative Agent Networks) in IBM's AI maturity model.
• Supports Bring-Your-Own-Model (BYOM) via provider CLIs, using existing subscriptions.
• Features extensive cost-optimization tools for token usage.
• Employs a 13-role, four-phase adversarial pipeline with mechanical enforcement.
• Prioritizes auditability through automated reviews and detailed logging.

Maintenance & Community

Roadmap items include planned roles like Sprint Analyzer and Security Auditor, alongside enhancements to context handoff and role transition hooks. Community links are not detailed, but related projects like CrewAI and BMAD are noted for their larger communities.

Licensing & Compatibility

Licensed under Apache 2.0, permitting commercial use and integration into closed-source projects.

Limitations & Caveats

Primarily software-engineering-focused. The L5 "software factory" vision remains theoretical. Cost tracking is planned but not implemented. Some LLM providers (e.g., Gemini 2.5 Flash) are incompatible. Multi-agent mode is relatively new (v0.4.0).