Token overhead in coding agents: the task used 0.67% but overhead used the rest (praveenvijayan.substack.com)

A recent analysis of coding agents revealed a staggering finding: only 0.67% of tokens were actually utilized for the primary task during a coding session, while a massive 99.33% were consumed by overhead. The agent was tasked with addressing a GitHub issue, which included reading, editing, testing, and committing code, but the bulk of the token use stemmed from re-reading tools, skill descriptions, and environmental data. This resulted in an overwhelming work-to-overhead ratio of roughly 1:150, highlighting a critical inefficiency in the use of tokens during agent operations. This finding is significant for the AI/ML community as it challenges the assumption that more extensive tasks naturally lead to increased token usage. The analysis suggests that for users operating coding agents regularly, understanding and optimizing the overhead costs linked to connected tools and repeated context is crucial. By measuring token usage effectively—through techniques such as monitoring the history of commands and selectively loading tools—the analysis provides clear guidelines on optimizing performance, reducing costs, and improving task efficiency. Ultimately, it underscores that improving context hygiene is more vital than simply expanding context windows, suggesting a reevaluation of how coding agents are configured for optimal performance.