Going Multi-Model and Multi-Tool

Claude Code is excellent, but it’s expensive. Ever since Opus 4.6 came out, it seems as though I use up my five-hour usage limit faster than I did before. There must be some unwritten change that bundles less usage with my pro plan now that the new models are available. Not to worry! I have low-tier ChatGPT and Gemini subscriptions, too! It’s time to try Codex and Antigravity.

One change in going multi-model is that I have to switch from CLAUDE.md to AGENTS.md — that was pretty straightforward, and I had Gemini help me with that. Unfortunately, I found in practice that Gemini is not great at actually following AGENTS.md. Because of this failure to follow saved instructions, I’ve instinctively used Gemini less than Codex.

GPT-5.3-Codex in Codex (actually VSCode Codex extension, because I’m on Linux) is better about following AGENTS.md than Gemini is, but I still feel as though I need to prompt it to review that file every time. But other than that, I absolutely love how inexpensive the model is. This week, I’ve been doing a big content push in System Design Practice, and most of the second half was generated by GPT-5.3-Codex, with medium reasoning effort. I have yet to do my own human review of the content, so I can’t guarantee the quality, but it sure is fast! People have said that “high” GPT-5.3-Codex is really good, and I’d like to try it, but right now, I’m treating Claude Opus 4.6 as my “staff engineer,” who takes care of major decisions. GPT-5.3-Codex is my content creation workhorse.

Claude is the best at following AGENTS.md — all I have to do is put something that says “read AGENTS.md” in CLAUDE.md, and Claude does the right thing. The “thoughts” that flashed on screen weren’t showing that Claude was reading AGENTS.md, so I asked directly. Evidence later indicated that specific directions were being followed: maybe I should have asked for a bowl of M&Ms with the brown ones removed.

Claude did seem to find some issues in the code generated by other models in OpenQuizzer. Claude did a major cleanup and improvement in code coverage by tests in both repos. The result is that I feel comfortable incorporating code from multiple models in multiple environments. The conventions are well documented, and if something breaks, a test is also likely to break.

One of my goals with this project has become to get as much as possible without intervening by hand. I say “Claude did seem to find some issues” because, as a hands-off manager, I just trusted Claude’s judgment. I didn’t verify that the issues were real. I feel as though having good documentation and good code coverage will help make that happen.