Metered: An Illustrated Field Guide to AI Agent and LLM Cost Optimization, with Prompt Caching, Context Engineering, Token Budgets, and FinOps for Engineers
English | October 4, 2026 | ASIN: B0HLXLF6G4 | 278 pages | Epub | 16.96 MB
Your AI agent works. Then the invoice arrives.Volume is on plan, customers are happy, and the model bill is three times the forecast. Nobody decided to spend it. The agent did, one step at a time.Metered is an illustrated field guide to AI agent and LLM cost optimization, written for the engineers who build agentic AI systems and are then asked to explain the bill. It shows where the tokens go and how to bring the cost down without making the product worse.What you will learnRead the bill. Build a token ledger that splits one task into tool definitions, system prompt, history and output, so you know which line to fix first.Prompt caching. Lay out a prompt so the cache is read, measure your cache hit rate, and see why a cache that is switched on can cost more than no cache.Provider pricing. Batch, flex and priority tiers, provisioned capacity, tokenizer changes and regional premiums on Amazon Bedrock, OpenAI, Google Gemini, Anthropic Claude and Azure.Context engineering. Trim tool definitions and MCP tool lists, shrink tool results, move instructions into skills, and decide when sub-agents save tokens.Harness engineering. Compaction, waiting without tokens and retry budgets in the agent loop.Token budgets and guardrails. Step, token and time limits, loop detection, AI gateways with a spending cap per customer, and a policy for coding agents such as Claude Code, GitHub Copilot and Cursor.FinOps for AI. Cost attribution with OpenTelemetry, cost per successful task, unit economics, pricing an AI feature and a forecast that finance can defend.Built to be used16 chapters, each with a running story told in comic strips, a worked example with full arithmetic and a one-page field card16 case files from public sources, covering Uber, GitHub, DoorDash, Shopify, Cloudflare, Cursor, Intercom, Manus, Writer and others18 hands-on labs with tables for your own resultsMore than 120 hand-drawn illustrations, plus checklists, worksheets, self-checks and Python listingsA 30-day plan and a one-page monthly brief for financeThe story that runs through itTallybird launches an AI support agent, and three weeks later the invoice is three times the forecast. Chapter by chapter, its engineers find where the money goes: a clock in the system prompt that defeats the cache, 32 tool definitions when a ticket uses five, a log excerpt resent twelve times, and a ticket that loops for 400 steps overnight. By the last chapter a resolved ticket costs a fifth of what it did at launch, and finance has a forecast with a range and a reason for each end of it.Who it is forSoftware engineers, AI engineers, architects, platform and LLMOps teams, and engineering leaders who run LLM applications in production. You need to be comfortable reading Python. You do not need a finance background.A note on providersThe worked examples use one provider's API so that a single story can run from the first chapter to the last. A guide at the front shows how to apply every chapter on OpenAI, Gemini, Bedrock and Azure, and an appendix compares their caching rules, tiers and usage fields side by side.
Links are Interchangeable - Single Extraction
