In short we clone the repo, build the docker container, and run agent in the container. The reason we do it this, rather than provide a docker image to the clearml-agent is two fold:

We actively develop our custom networks and architectures within a containerised env to make it easy for engineers to have a quick dev cycle for new models. (same repo is cloned and we build the docker container to work inside) We use the same repo to serve models on our backend (in a slightly different container)
I guess we could build and push the container to docker and reference that image in the clearml-agent. What do you think about this workflow? They are the main reasons for not using the startup_bash_script as provided