julep

Why we included this project

Julep tackles a pain most agent frameworks ignore: what happens when a long multi-step run dies halfway. It models agents as durable dataflows, so you write the flow in ordinary Python and the graph of tool calls, retries, timeouts, and branches compiles into a frozen wire format. If a run crashes, it resumes instead of restarting; each step can be explained through a derived projection, and the runtime refuses to let the model call any tool it was not explicitly allowed to use. Teams building production workflows with decision loops, parallel branches, or long-lived runs will find the optional Temporal layer useful for durable execution and retries. The quickstart runs without an API key and both Python and Node.js are supported, so a small team can test it against its own stateful agent scenario before committing.