harvard-cns/orla

Separates workflow logic from execution by decomposing agentic systems into stages, then intelligently routes them across heterogeneous models and backends via a Stage Mapper. Includes a Workflow Orchestrator for scheduling and coordinating stage execution alongside a Memory Manager that optimizes KV cache sharing across inference requests. Provides both a daemon service and Python SDK for building multi-agent systems that combine LLM calls, tool invocations, and distributed infrastructure.

208 stars.