Skip to content

crewmaster

The Ultimate Framework for Building AI Teams 🚀¤

CrewMaster v2.0.0 is a framework-agnostic Python library for building, orchestrating, and evaluating type-safe multi-agent systems. It lets you define Operations, wire them into DAGs, and execute them with structured, streaming output — all without being locked into a single LLM framework.

Autonomous agents + developer ergonomics¤

CrewMaster is designed for developers who want to:

  • Define Operations — typed, recursive units of work with declarative produces/consumes.
  • Configure Agents with identity prompts, default tool scopes, and context projections — no framework coupling.
  • Build ExecutionPlans — DAGs resolved automatically from Operation trees, with dependency ordering and runtime assignments.
  • Share Tools across agents via a ToolRegistry with semantic retrieval at runtime (request_capability).
  • Write prompts using Jinja2 blocks (blocks://) stored in a BlockStore, composed and rendered by the PromptEngine.
  • Execute operations synchronously with execute() or stream with execute_stream(), yielding typed events.
  • Evaluate agent outputs with snapshots, datasets, criteria, and evaluation reports built on the same Operation primitives.

Why Choose CrewMaster?¤

  • Framework-Agnostic: Pluggable runtime drivers — PydanticAIDriver and LangChainDriver out of the box. Bring your own by implementing the RuntimeDriver protocol.

  • Type-Safe by Design: Every Operation declares what it produces and consumes. Pydantic models enforce contracts between nodes. mypy and IDE autocompletion work everywhere.

  • DAG-Based Execution: Multi-node pipelines with typed data flow. Topological ordering, retry with secure_retry_execution, and progressive tool disclosure via request_capability.

  • Collaboration Primitives: Built-in protocols for debate, consensus, supervisor, blackboard, and GrillMe — compose multi-agent reasoning without boilerplate.

  • Conversation Abstractions: dialogue() for human-agent interaction with structured clarification, and channel_dispatch() for ordered multi-agent chat with history accumulation.

  • Built for Production: Evaluation framework with snapshots and datasets, streaming with SSE, and a sandbox UI for interactive testing. Designed for CI/CD integration from day one.