Stop Burning Tokens: Cost Engineering for AI Coding Agents

₹2,100

Your AI coding agent is worth its cost. But how much of your bill is pure waste? Most of it comes from the same few causes: stale context, oversized instruction files, premium models on easy tasks, broken caches and unattended loops. None of them needs a new tool to fix. They need the discipline you already apply to latency and memory: measure, find the hot path, optimise it. This 48-page guide turns that into a routine you can start today, with a 30-day rollout and a one-page checklist. What’s inside - Why agent bills explode: how context is re-sent on every request - Measure first: /usage, /context, and the four numbers worth tracking - Model routing: mid tier by default, premium for planning, cheap subagents for legwork - The context-file tax: keeping CLAUDE.md and AGENTS.md lean - Context-window discipline: clear, compact, filter noisy output, fence off files - Prompt caching: prices, lifetimes and the habits that break it - Taming tools, MCP servers, subagents and agent teams - Local and hybrid fallback with Ollama, LM Studio and llama.cpp - Guardrails against bill shock: budgets, turn limits, spend limits and alerts - The cost-engineering checklist and a 30-day rollout - Glossary and resources Every chapter has a Key Points summary, and most include a worked example you can adapt: a usage logger, a cheap exploration subagent, a test-output hook, a cache-enabled API call, a local-model launcher and a budget-capped CI step. Who it’s for: developers and teams who use Claude Code, Cursor, Codex or similar agents daily and want the same output for less spend. Prices and features were checked in September 2026. Worked examples use illustrative numbers unless a source is cited.