How AI coding agents rack up costs through small, reasonable-seeming decisions

Agentic coding tools can become surprisingly expensive not through obvious mistakes but through many individually justifiable micro-decisions that accumulate over time. A developer analyzing their own billing identified five key cost drivers: keeping too many files in the main agent's context, reading more files than necessary, using powerful frontier models for simple lookup tasks, generating long outputs inline instead of delegating to cheaper sub-agents, and allowing context bloat that silently inflates costs on every subsequent turn. Because each decision feels harmless in the moment, the author argues that personal discipline alone cannot solve the problem. Instead, they built a structural solution comprising a per-tool-call spending hook, a model router, and a parent-child agent delegation system. Notably, the cost-cutting measures also improved output quality by keeping the main agent more focused and accurate.