Yeah the ultimate goal I'm trying to achieve is to flexibly running tasks for example before running, could have a claim saying how many resources I can and the agent will run as soon as it find there are enough resources

Checkout Task.execute_remotely() you can push it anywhere in your code, when execution get to it, If you are running without an agent it will stop the process and re-enqueue it to be executed remotely, on the remote machine the call itself becomes a noop,

I can comment it on the github issue

Yes please do 🙂

How can I do to help extend it?

How about a CLI tool, like what we have with "clearml-task" ?

This is so awesome

Thank you ! 😊