Built this with Cursor: an autonomous coding agent that plans and writes code like Devin-style agents do, but stops for approval before executing anything risky, instead of running fully unsupervised. Paired it with a standalone QA agent whose only job is reviewing what the coding agent produces before it ships.
Goal isn’t to be “another AutoGPT clone” — it’s to make autonomous coding agents something you can actually trust enough to use daily, without babysitting every line.
Going open-source soon. Curious what this community thinks the biggest risk is with full autonomy in coding agents — is approval-gating enough, or is there more needed?