Julep

You will want this when the thing you are automating takes hours rather than seconds. Flows are ordinary functions with decorators on them, and a run that dies halfway comes back at the step it died on instead of at the beginning. Resumability is the trait you feel every single day, and it is the reason to look here at all.

It is not a coding agent. No shell, no git, no file edits, so do not reach for it to refactor a repository. Pick it for long-running data and agent pipelines in Python. Pick LangGraph if you want a bigger community and more examples to copy.

The risk is maturity. The published install command asks for a pre-release package, and the project labels itself preview. Building a production pipeline against an interface that is still allowed to change means the migration bill arrives on a schedule the maintainers set. Nothing about durable execution helps when the decorator signature moves under you.

Pin an exact version and read the changelog before every upgrade. The one thing done right: secrets sit in an encrypted vault inside a control plane you host yourself, rather than in a vendor database you cannot inspect.

1. Authoring is by decorator, and the decorated flow compiles to a frozen intermediate representation rather than being interpreted at call time. 2. That representation is what executes, so one definition runs in memory during development and on Temporal or DBOS in production. 3. Retries are described as side-effect safe, which requires the runtime to record which effects already committed.

No benchmark accompanies the project, so capability is asserted rather than measured. The design should outlast model changes, because durability lives below the model call rather than inside the prompt.

The product exists and the business does not yet. Roughly 6,600 stars of developer attention, a first release in April 2024, and no paid tier published, so revenue has to arrive later from hosting or support. That sequencing is normal, and it also means the next eighteen months depend on a fundraise rather than on renewals.

Moat: durable execution is a crowded idea with funded incumbents already selling it as infrastructure. Likely acquirer: an orchestration vendor buying an agent-shaped front end, or quiet absorption into a larger Python stack. Position: build with it, do not build a roadmap on it.

The demo is a pipeline surviving a restart. Procurement is short because there is nothing to purchase: sixty engineers cost zero beyond model spend, which is the meter nobody puts in the budget. It runs unattended, so it becomes a step in the pipeline we already operate without standing up a new runner.

What is absent is everything security asks for: no single sign-on, no audit trail, no retention statement, no named support channel. Onboarding is the other cost, since this is Python only and the decorator model takes a mid-level engineer several days. Approved with conditions: internal pipelines only.

Apache-2.0 and pure Python, which means the whole thing is legible and a fork is one clone away. It speaks MCP as a client, so external tools arrive over a protocol instead of through bespoke plugin classes, and the CLI runs and ships flows without a dashboard sitting in the loop. I point it at whichever provider key I feel like using.

The hole is offline. Local models are not supported, so the box under my desk sits idle while this talks to somebody else's endpoint. For a project that markets self-hosting that is an odd gap. Grudging respect for the code, a shrug for the model layer.