AI ยท
Why AI coding agents need a new measure of value
As agentic coding workflows consume far more compute than simple chat turns, some argue that token counts are the wrong yardstick for judging real productivity.
The way developers evaluate AI assistance may be due for an update. When an AI agent takes on a coding assignment, it does much more than produce a single reply. The system lays out a plan, writes code, executes tests, and cycles through corrections when something fails. This back-and-forth can make one request very expensive to process.
Compared with a basic chat prompt, an agentic coding session can consume 10 to 50 times as many tokens. That gap grows as the tool gets better at handling complex tasks, because each improvement tends to trigger additional rounds of execution, checking, and revision.
Against that backdrop, focusing only on how many tokens are used misses the point. The more meaningful question is what actually gets finished and shipped. A coding agent that burns more tokens but delivers a working feature may be far more valuable than a lean prompt that returns a snippet requiring hours of human cleanup.
Organizations are beginning to ask how to price, monitor, and compare these systems when raw token volume no longer lines up with user benefit. The shift could reshape how AI coding tools are built, sold, and measured inside engineering teams.
Source: MIT Technology Review
Written with AI from the linked source