Note that Claude overuses tokens. I guess they get paid by the token. It is cliched in this way, sort of like if a programmer were to get paid by the line. Other LLMs can produce clean results with far fewer tokens than Claude.

I believe we already have the correct word (agent).

Agentic coding.

To me this implies that it's more agent than person... like when you use claude in github actions vs using it in your ide.

Agentic coding works via the IDE, GitHub Issues, and command-line tools like OpenAI Codex and Claude CLI. Agentic coding is AI-based feedback-driven multi-step coding. In general, an agent implies that the agent is working for someone, which in this case is a person.