Match a Prefect replacement to what your flows do: scheduled Python data pipelines, container jobs on Kubernetes, or durable execution for apps and AI agents that must survive a crash. Each of those has its own specialists, so start from the job, not the brand, and move one real pipeline across before you commit the rest.
Prefect itself has changed what it is. Its homepage now sells workflow orchestration for data, AI agents and MCP servers, and its menu lists FastMCP, MCP hosting and an MCP gateway next to the orchestrator. If you came to it for plain Python pipelines, that drift is a fair reason to look around. If you came for agent work, it may be pulling in your direction.
The rival that isn't a rival any more
The first name most people put on a Prefect shortlist is Dagster, for its asset-based way of thinking about pipelines. Dagster's own site now says Dagster is part of Prefect. So if your reason for leaving is the company, the strategy or the contract, moving there may simply bring you back to the same owner. That's why it isn't in the ten below.
The wider lesson for this category: check who owns the thing before you spend a month migrating to it. Orchestrators sit underneath everything, and they're the last piece of infrastructure you want to move twice.
Three jobs an orchestrator can do
Most confusion in this category comes from lumping three different kinds of tool together. Here's how the ten split. The kind matters more than the name.
| Job | What it looks like | Tools on this page |
|---|---|---|
| Python data pipelines | Scheduled batch jobs, dependencies between tasks, backfills | Apache Airflow, Kestra, Luigi, Metaflow, Windmill |
| Kubernetes and ML runtimes | Containers on a cluster, parallel jobs, training and serving | Argo Workflows, Flyte |
| Durable execution for apps and agents | Code that resumes after a failure, queues, long-running agent steps | Hatchet, Inngest, Temporal |
Some tools stretch across lines: the declarative one handles infrastructure and AI work too, and the script runner happily runs SQL, Bash and TypeScript next to Python. But the row a tool starts in tells you what it does best.
Ten orchestrators to consider
Each gets a one-line description. The choosing happens in the sections that follow.
Apache Airflow
The community-built incumbent for writing, scheduling and monitoring workflows as Python code, with pipelines that can be generated dynamically.
Argo Workflows
A container-native engine that runs parallel jobs on Kubernetes, described either as DAGs or as step-by-step sequences.
Flyte
An open-source runtime for training, serving and agent workloads that is built to pick up again after a failure.
Hatchet
Durable workflows, background tasks and AI agents on one platform, using checkpoints so a failed run picks up where it stopped.
Inngest
Reliable functions, queues, crons and agents without running extra infrastructure, using checkpoints inside ordinary application code.
Kestra
Declarative orchestration where workflows are written as configuration rather than code, spanning data, infrastructure and AI jobs.
Luigi
A small Python package for chaining batch jobs, handling dependency resolution and failures, with a visualiser for the resulting graph.
Metaflow
Pitched at data scientists who want to develop and debug a flow locally and then run the same code in production unchanged.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
Temporal
An open-source durable execution platform, saving workflow state at every step so code carries on after crashes and restarts.
Windmill
A self-hostable platform for running scripts and flows in Python, TypeScript, Go, PHP, Bash or SQL, open-source and easy to host on your own kit.
Self-run or hosted, and what that costs
This is a category where the open-source edition is the norm. Six of the ten show no hosted price on their own sites at all: they're self-run projects, or the paid version is quote-only. The other four publish numbers, and on their pricing pages as checked on 13 November 2026 the shapes differ. Where a plan charges per million runs or actions, the rate per million is on the pricing page. Flat team plans, where offered, are set out on the same pages.
So the honest comparison isn't licence against licence. It's the cost of someone running the scheduler, the database and the workers yourself, against a bill that grows with the number of runs. Teams that already run Kubernetes well usually find self-run cheaper. Teams firing millions of short agent steps should model the usage bill on a real month of traffic before they sign anything.
A one-hour spike before any migration plan
Pick the pipeline that annoys you most in Prefect today, ideally one with a retry and an external call. Spin up the candidate locally; every one of the ten has an open-source edition you can run yourself. Port that single flow, keeping the logic the same.
Now break it on purpose. Kill the worker halfway through a run and see what happens when it comes back: does the run resume from the last good step, restart from scratch, or sit there looking healthy while doing nothing? Then trigger a backfill for last week and check the tool handles dates the way your data team expects. Those two results tell you more than any feature grid, because retries, resumes and backfills are where orchestrators earn their keep or fall apart.
If both pass, keep the old orchestrator running and move the next painful pipeline. Only switch Prefect off when nothing important depends on it.
Where Prefect migrations go wrong
Picking a durable execution engine to schedule nightly batch jobs, or a batch scheduler to run an agent that waits hours for a human. Both work on a demo and hurt in production. Match the row in the table first.
Rewriting everything in the new tool's idiom on day one. Port flows as they are, prove they run, then refactor. Doing both at once makes it impossible to tell whether a bug came from the move or the rewrite.
Forgetting secrets, retries and backfills until late. Check how each tool handles them before you commit, because those are the parts that hurt later. If you're comparing the big Python scheduler in more depth, my list of Airflow alternatives goes further, and if your workloads look more like long-running application code, these durable execution options cover that side.
Which kind to choose
Scheduled data work in Python with a data team that likes code: a Python pipeline scheduler, and the long-established community-built scheduler is a safe default. Workflows as configuration that non-Python colleagues can read: the declarative option. Already on Kubernetes, with training and serving jobs: a Kubernetes-native engine. Product features or AI agents that must not lose their place when something crashes: a durable execution platform, hosted if your team is small, self-run if you have the platform people to look after it.
Do you make workflow orchestration or durable execution software? Tool makers can ask to be considered for a featured place on this list.
Want your tool on this list? Featured spots and reviews here are paid partnerships. Here's how to partner with me.
