Meta-Models Releases Legal Agent Workshop Repository for Building AI Legal Assistants

Overview of the Legal Agent Workshop

A new open-source repository titled legal-agent-workshop, published by meta-models on GitHub, provides practical code and instructions for building an AI legal agent. The project utilizes Muse Code and Muse Spark tools to guide participants through a structured two-session workshop format focused on developing, testing, and verifying an AI-powered legal assistant.

According to the repository documentation, the workshop is divided into distinct phases designed to move from basic agent construction to implementing robust verification and safety checks.

Session 1: Building the General Legal Agent

During the first working session, participants clone the repository and set up a workspace restricted strictly to the `agent/` directory, keeping case files and world simulations outside the primary workspace.

• Agentic Loop: Built using plain OpenAI Python SDK code.
• Pre-Built Kit: Includes modular files for individual tools, document reading via the Files API, and a case-management system featuring quote-checked facts.
• Workflow: The agent learns to draft a checklist, plan its approach, investigate materials, and conclude with an unverified preliminary memo.

Session 2: Establishing Trust and Verification

The second session focuses on addressing the limitations and unverified nature of the initial output by introducing a controlled verification layer.

• Review: Participants first evaluate the Session 1 memo and checklist against known omissions using helper verification commands.
• Verification Layer: Developers build two primary checks running outside the agent's control—validating the checklist against case ground truth and enforcing memo rules.
• Advanced Controls: Utilizing branching features, the workshop demonstrates handling refusals, escalations, the release of new evidence, session resumption, and formal attorney sign-offs.

Repository Structure and Monitoring

The project repository is organized into specific directories to separate concerns between agent logic, case data, and user interfaces:

• `agent/`: Starter code and the primary workspace for Muse Code.
• `cases/`: Contains case files, such as `northbridge-v-caldera`, consisting of objectives and documents.
• `ground_truth/`: Stores verification benchmarks like `northbridge-v-caldera.json`, which remain strictly hidden from the agent.
• `world/`: Simulates the outside world with withheld follow-up documents and release scripts.
• `ui/`: Features a pre-built Agent Console (`python3 ui/ui.py` running locally on `http://127.0.0.1:8765`) to monitor runs, track checklist verification, and manage evidence release.
• `HANDBOOK.md`: Contains participant instructions, setup guides, and required ground-rules prompts (section 1.7).

Practical Implications and Verification

This workshop provides developers with a hands-on framework for understanding both the capabilities and the inherent reliability challenges of generative AI in legal contexts. By separating core agent logic from ground-truth verification layers, the repository highlights methods for maintaining oversight in automated legal workflows.

Interested developers can review the complete source code, repository structure, and instructions directly on the GitHub repository. Verification of updates and instructions should be conducted by reviewing the official repository README and HANDBOOK.md files.