Design Cost Ledgers for AI Coding Agents

This technical guide explains how engineering teams should record and analyze AI coding-agent costs at the session level. It defines an "AI coding agent cost ledger" as an append-only, session-centered record of model requests, tool calls, file reads, retries, approvals, verification events, and estimated provider costs. The article recommends tracking five key metrics (total session cost, repeated input ratio, cost per accepted change, idle approval time, verification coverage), attaching every cost event to a session_id, using append-only event records, and rolling out a simple Postgres schema. It also covers implementation patterns (request wrapper, purpose labels), alerting rules, privacy/security practices, and a staged rollout plan to convert ledger data into product decisions like model routing, budgets, and workflow improvements.