# Multi-Agent Workflow Lab

An open-source framework for testing and observing multi-agent delegation, tool execution, MCP workflows and permissions, with policy checks, replayable runs and diagrams of who delegated what to whom.

Multi-Agent Workflow Lab, or MAWL, is a testing and observability framework built for questions a model's final answer cannot explain: which agent delegated a task and to whom, what permissions and budget the child received, whether an agent tried to escalate its privileges or loop, and whether a run can be inspected, compared or replayed without repeating its side effects.

It models delegation as parent and child tasks with capability checks, depth and fan-out limits and loop detection; tool execution through a registry with schema validation, allowlists, policy and optional approval; MCP access through official stdio and streaming HTTP client adapters plus an in-memory mock, treating MCP data as untrusted; and prompts as versioned, hashed assets with provenance tracking. Runs are recorded as append-only events in SQLite or JSONL, with structured logs, task traces, delegation graphs and budgets.

It is an experimental release, described as suitable for local development and policy testing rather than production infrastructure.

Website: https://github.com/christiangrey922/multi-agent-workflow-lab