OpenRath swaps chat logs for tensor-like agent state

What it does

OpenRath is a Python framework that models multi-agent workflows as explicit, composable objects. Session carries runtime state and collaboration lineage; Agent acts as a reusable transformation layer on that session; Workflow nests these layers into larger systems; and Selector routes between them at runtime. The design explicitly mirrors PyTorch concepts—Session is a Tensor, Agent is an nn.Linear layer, Sandbox is a Device, Memory is a Parameter—so you build agent clusters by composing dataflow rather than piping chat strings.

The interesting bit

The framework treats Session as the single flowing runtime value rather than an agent-owned message history, which means you can fork, merge, detach, and trace branches across a cluster of agents without losing provenance. This shifts the problem from prompting a model to routing and transforming state—a distinction that matters when you stop building chatbots and start building agent clusters.

Key highlights

• Session is a first-class, forkable and mergeable data structure with lineage tracking, not just a conversation transcript.
• Agent, Tool, Memory, Sandbox, and Workflow are decoupled primitives that plug together like neural-net layers.
• Memory persists across runs via local BM25 recall (with optional embeddings) or external backends such as OpenViking.
• Sandbox placement determines where tools execute—local process or containerized OpenSandbox—without hardcoding execution to the host.
• Selector provides LLM-backed runtime routing between self-describing workflows, keeping if/while logic in plain Python.

Caveats

• The README targets massive agent clusters, but evidence of production hardening at that scale isn’t visible in the sources.
• Key backends like OpenSandbox and OpenViking are optional or external, so out-of-the-box coverage depends on which integrations you bring.

Verdict

Worth a look if you’re building non-trivial agent orchestration where traceability, branching, and memory matter more than a single chat loop. Skip it if you just need a thin wrapper around one model call.