has anyone been experimenting with fully-agentic SWE workflows?

I'm really enjoying using Clod to make iOS apps, the problem is I'm a data professional by trade, which means I know my way around SQL, Python, and… um… YAML? 😅 But hey, I also know my way around github-based development, CI/CD pipelines, and automated merge gates, so let's see how far that gets me. Here's what I'm using. It's not great , but it's a big step up from a single Clod sesh just trying its best: I give a spec to the PM (Claude Fable), which writes a plan of work The plan reviewer (GPT Astra) reviews it from a "is the the right approach?" perspective No back-and-forth here, if ChatGPT doesn't rubber stamp it then it's raised to me to decide. Once the plan is good, the PM gives it to an engineer (Claude Opus), which works until a PR is ready and all tests pass A PR reviewer (GPT Sol) asks "does the code do what the description says?" Failing reviews are given to the PM at first, not the engineer. The PM directs the engineer in fixing any issues Steps 5, 6, and 7 should only loop 3x before the PM surfaces the PR to me and I figure out if the blockers are dumb nits or not PRs automerge once there's a passing review Implementation details: the 3 round limit needs actual enforcement, if it's just a prompt the fable PM will happily let it go to 12 rounds >.< autocompact ceiling set to 400k tokens for all agents Claude Code is on the max plan, Codex is on the #20 sub I am quite sure others are doing this much more sophisticatedly, here's hoping for an educational roast!