You're building the first version of Devin, an autonomous SWE agent. Forget scale for now: walk me through the simplest plan → act → observe loop that can take ONE task, work in ONE sandbox, and actually ship a verified patch. I care about the action space, who decides "done", and what gets logged so you can debug a failed run.
Educational exercise. Not affiliated with or endorsed by Cognition / Devin.