4 sessions on paperbanana with a coding agent: the agent failed at least once in 4, and 0 finished without a human stepping in.
Takeaway: Next time I'd check upfront whether the sandbox even allows running python or writing files before asking the agent to run a whole pipeline script, since it burned a lot of turns just probing what was blocked.
Takeaway: Next time I'd check upfront whether the sandbox even allows Bash execution of scripts that hit external APIs, since none of the three approaches (heredoc, /dev/stdin, direct venv binary) got past the permission wall.
Takeaway: Next time I'd start by manually trimming what context gets loaded, since the agent kept blowing past the 128k token limit by dragging in the whole paper, config, and image files at once.
Takeaway: Next time I'd tell it upfront which python environment has matplotlib and paperbanana installed, since it burned a lot of turns discovering that the system python, the project venv, and pip were all broken before it found the paperbanana venv that actually worked.