Cursor published an article yesterday - "Scaling long-running autonomous coding" to explore how to enable intelligent agents to work continuously like engineers on a project for an extended period.

@wilsonzlin used a planner agent for management and division of labor, and worker agents for programming to complete tasks, allowing AI to work nearly a week straight, writing over 1,000 files and 1 million lines of code, resulting in a web browser.

Using GPT-5.2 Pro for estimation, each line of code requires multiple rounds of reading context, trying/rolling back/fixing, with different agents planning, executing, and evaluating. Each line of code may consume 2-5k tokens. The entire project approaches 2 to 5 billion tokens; for simplicity, roughly estimated at 3 billion tokens.

Assuming:
90% input tokens
10% output / reasoning tokens

Using GPT-5.2 standard pricing:
Input: $1.75 / 1M tokens
Output: $14.00 / 1M tokens

The final cost is approximately $8,900. If not based on workload and hours, "hiring" such an endless AI engineer would cost $35,600 per month, and $427,000 per year. This is comparable to the cost of hiring an L4 engineer at Google.