Claude Code is usually described as Anthropic's coding agent. It is more useful to think of it as a harness. A loop prompts a model, runs tools, checks results, and decides when to stop, and the model is a swappable part. Once you see it that way, interesting doors open. You can point the same harness at any model on OpenRouter, run a free open-weight model for prose and a frontier model for code, and find out which parts of "agentic coding" were the model and which were the loop.

This session shows how to do that in both Claude Code and the Codex CLI, since the same trick works in both harnesses, and then spends most of its time on what happens next. The loop is the bill, so I'll show how a free model cost me nineteen dollars in an afternoon. Cheaper models fail in different places than expensive ones, so I'll show a text-only model that reports "model not found" when you paste a screenshot, and what it takes to notice. Verification is the part you cannot outsource, so we'll look at the ways an agent quietly moves the goalposts and the guardrails that stop it. We'll finish with the loop ladder, from an interactive session through headless runs to hosted agents, and which rung each kind of work belongs on.

You'll leave with a working multi-model Claude Code configuration, a cost model that predicts the bill before you run the loop, and a short list of what to verify no matter which model did the work.

50-minute talk. Intermediate to advanced. Live demos with a recorded fallback. Material drawn from my forthcoming O'Reilly book, Claude Code: Up and Running.