OpenRath is a production-grade multi-agent workflow orchestration framework with a PyTorch-like composable API. It provides a structured session abstraction (dataflow with lineage, fork/merge, placement via `.to()`), pluggable LLM providers (OpenAI, Anthropic, LiteLLM), sandboxed execution (local subprocesses, containerized OpenSandbox), persistent memory (local filesystem with BM25, OpenViking vector backend), and a durable runtime with effect ledger, checkpoints, interrupts, and lease-based worker fencing. The framework includes a full AgentServer (REST + SSE) with authentication, authorization, audit, and an OpenAPI contract. It supports defining agents and workflows as composable Python objects with `@step`/`@router` decorators and a compiler that validates resource manifests. The codebase is heavily tested with conformance, integration, and benchmark suites, and follows modern Python patterns (protocols, dataclasses, async/await, asyncio runtime bridging). While the repository contains evaluation and memory backends that touch on data aspects, its dominant character is a software application framework for building and deploying AI agents.