Skip to content
MAESTRO

What is CARL?

CARL is the format reasoning chains are written in. A chain is a list of typed steps with dependencies, per-step context queries, and execution settings. The reference implementation of the format is the mmar-carl Python library (Collaborative Agent Reasoning Library): it reads a chain, builds a directed acyclic graph (DAG), runs everything that can run in parallel, and pulls relevant context from your input for each step.

In MAESTRO, MAGE writes chains in this format and the runtime executes them. CARL chains can also be authored and run directly from Python.

Key features

Section titled “Key features”
  • DAG-based execution — steps parallelise automatically based on dependencies.
  • RAG-like context extraction — substring or FAISS vector search pulls relevant context per step.
  • Many step types — LLM, Tool, MCP, Memory, Transform, Conditional, Structured Output, AgentSkill, and multi-agent orchestration (handoff / supervisor / debate / parallel sampling / human-in-the-loop).
  • Async + streaming — execute_async and stream_async with per-step callbacks.
  • Evolution & evaluation — genetic search over chains, metrics, dataset evaluation, reflection.
  • Observability — execution traces, Gantt charts, token/cost breakdowns, Langfuse.
  • OpenAI-compatible — OpenRouter, Azure, Ollama, vLLM, LM Studio; plus a native Anthropic client.

Installation extras

Section titled “Installation extras”
Terminal window
pip install mmar-carl                 # core (substring search)
pip install 'mmar-carl[vector-search]' # FAISS semantic search
pip install 'mmar-carl[openai]'        # OpenAI-compatible providers
pip install 'mmar-carl[mcp]'           # Model Context Protocol
pip install 'mmar-carl[skills]'        # AgentSkills (+ pdf + pptx)
pip install 'mmar-carl[viz]'           # PNG chart output
pip install 'mmar-carl[all]'           # everything

Requires Python 3.12+.

Where to go next

Section titled “Where to go next”