langcontinuation: durable asynchronous LLM workflows

Want to save 50% on your LLM bills?

Me too.

Want to avoid writing painful workflows on batch processing? Especially with fork/join semantics or variable-length flows?

Me too.

I built langcontinuation to address both of these needs in one hit.

It’s a langchain-inspired project that takes advantage of some old computer science that formalized the concept of goto; appropriate considering most workflows are little more than a dag of goto statements.

The live executor is suitable for prototyping in the current form. You can run workflows that don’t require a human-in-the-loop and use the included Anthropic provider.

The batch executor is where this project shines. It includes full provenance tracking and logging of all material transitions; an included example html debugger generates standalone, interactive pages to see the status of a flow.

What’s the catch? It’s more like a lego set than a working workflow engine. You have to put some parts together to submit workflows, keep them running, and tend to human-in-the-loop semantics.

You also are bound to Anthropic because OpenAI doesn’t have a decent Rust client and my personal time is guarded enough I don’t have a time to write another LLM client library (see: claudius)