Controlling AI Coding Agent Costs: Budget Management for Long-Running Jobs

AI coding agents are powerful, but they burn through API budgets quickly. An overnight job that reads a large codebase, makes edits, runs tests, and self-corrects can cost £20-40 in tokens. Run that daily for a month, and you're looking at £600+. The key difference from chat APIs: you're not paying for conversation turns, you're paying for computational work. More iterations, more cost. Silence detection is critical-if Claude Code enters a loop trying the same fix repeatedly, it'll cost £40 instead of £8. Scoping tightly reduces iterations by 40-60%: instead of "refactor the auth module," say "extract password hashing logic from auth.ts into utils/hash.ts, update imports." Pre-warming context with CLAUDE.md saves 20-30% in tokens by letting Claude Code skip exploratory reads. Running 10 small jobs instead of one large job saves tokens because each starts fresh with a smaller codebase context. Self-correction is expensive but worthwhile for complex tasks-disabled for simple ones, enabled where failure risk is high.