The idea behind the cartoon
Every “AI can code now” conversation I have eventually collapses into the same confusion: chat-based AI and agentic AI get talked about as one category, when the actual work of using them looks nothing alike.
Chat is prompt-by-prompt. You say “change the font,” it changes the font, then it waits. A finished, working website has to be assembled by you, one instruction at a time, which is why the top panel is a boss buried in sticky notes handing off micro-tasks nobody upstream ever wrote down as a plan.
Agentic coding starts from the outcome instead. You describe what “done” looks like, and the system runs its own loop: understand the requirement, inspect what already exists, search for how to do it correctly, implement, test, fix what breaks, deploy, verify, all without you supplying every intermediate step. Anthropic describes Claude Code’s loop in exactly those terms, and OpenAI positions Codex the same way, as software that completes engineering work end-to-end rather than answering one prompt at a time.
The model did not get smarter between these two panels. The unit of instruction changed from a task to an outcome, and that is the difference this cartoon is actually about.
