A while back Prashant Anand asked me about what my agentic coding workflow looked like. I procrastinated a while on a...
A while back Prashant Anand asked me about what my agentic coding workflow looked like. I procrastinated a while on a write-up since everything's in such a state of flux, but actually sitting down and trying to describe just my current "snapshot" was still a pretty useful exercise. A few things I'll highlight:- I wrote this up after having given Opus 4.6 and GPT-5.3-Codex (xhigh) pretty decent shakes (billions of cached tokens on a few different projects). I'd rank GPT-5.3-Codex as the best coder for my workload atm, with Opus 4.6 and GPT-5.2 as my best planners and reviewers. (I haven't tried Gemini 3.1 for real work, but Gemini 3 still had instruction following issues that disqualified it from touching my code, so I'll probably skip 3.1 as well).- I've settled into a basic spec-driven loop that works for me, described pretty faithfully in Prashant's newsletter. Besides front-loading with multiple iterative/sequential passes of PLAN/IMPLEMENTATION, I think the most unique/impactful technique for me is that I've found having *independent* (separate session) review passes from multiple models (Opus 4.6, GPT-5.2, GPT-5.3-Codex) and passing feedback back to the coder model for remediation was surprisingly useful, catching way more issues than any individual model/pass would.- Maybe not emphasized enough, but I'm also all about the minimum amount of scaffolding to get my work done, and I'd encourage resisting yak-shaving as much as possible since everything is changing so fast. It's tough to say that anything you spend a lot of time tweaking now will be relevant/useful in 6 months.