Agentic development isn't free. A subscription hides what it costs: you pay the same every month, however much work the agents do, and an API bill only shows one total. Without knowing what a change costs, you can't make good choices about which model builds it, how deep the review goes, or how many builds you run at once.
The first article is about measuring it: tying every token to the pull request it was spent on. The second is about what to do when an engineer keeps hitting the subscription's limit.