Most of the token cost in agentic coding is not the code generation itself. It is everything around it: re-reading the same files, agents rediscovering context they already had, fan-out where every task spawns new tasks, and the expensive model being the default for every step.
I learned this the embarrassing way, burning my Claude Max limit in 30 minutes while researching how to save tokens. So I did the research properly. The write-up reached #2 on Hacker News with 212 comments, and everything useful from it went into an open catalogue: cost tracking tools, caching and routing patterns, context management practices, research, benchmarks. 200+ entries, each with a primary source and a verification date.
The agentic coding section is probably the most relevant for this forum: tools that show what your agent actually spends, and patterns for keeping context small without making the agent dumber.
I work at Quesma. The list is CC-BY, corrections and additions welcome. Curious what actually moved the needle for people here: caching, routing cheaper models, or just keeping context shorter?