Build the factory, not the car · AI maxing lessons

$1.2k / month on AI coding tools. Here's what our AI maxing dev workflow actually looks like. Philip and I are building Oxygen, an AI CRO agent. Last month we went full throttle · token throughput at the ceiling, 32-day GitHub streak, vibe coding the whole stack. The lesson: build the factory, not the car. Tesla doesn't ship cars, they ship the line that ships cars. What that means in practice: 1) Pick the right model. Codex is the autistic one. Reliable, methodical, brutal at backend. Claude Code is the ADHD one. Cracked on frontend and design. Stack: 2x Codex 20x Max + 1x Claude Code 20x Max. We usually have 4-8 agents running in parallel. 2) Wire MCPs for observability. The dev tools your agent actually needs: database, LangFuse for agent tracing, Vercel for deployment. The real unlock is the agent can read logs, errors, and traces. Bug patching becomes 10x better. Without it, the agent is flying blind. 3) Skills with real corpus. Build skills that load your top 20 posts, the slop blacklist, the hook patterns. The agent matches your actual voice instead of inventing one. 4) Fight the AI in plan mode before any non-trivial change. The spec is a back-and-forth discussion. You push back, the model pushes back, you converge. Skip this and you get slop on pass 1, then 5 iterations cleaning up. Do this and the change works on attempt 1. For bigger changes, the spec matters more. 5) Modular by default. No 10k-line files. Token throughput is now so high it's literally impossible to re-read every diff. The only way to stay safe is tight modules, tight contracts. Agents context-rot on monoliths. 6) Greenfield: let it rip. Brownfield: think before you token. First month felt unstoppable. Then last week the bill came due: context rot, duplicate utilities, 3 architecture pivots leaving deprecated code under the new layer. The dopamine of AI maxing is real. The cleanup tax is also real. Curious what others are running.