Coding agent cost review, starting with one team

Everything is written down and stays with your team, whether or not you work with us afterwards.

• Spend by tool, model and type of task for the team in the review.
• The changes made with the team: default models, routing, context settings and budgets.
• Cost per merged change before and after, with review effort and rework checked on the same kind of work.
• A plan for the next teams: what to roll out, in which order, and what to measure.

In the first week we read the usage and billing exports, talk to the engineers about how they work, and replay a sample of changes the team has already merged. That gives the baseline: what a merged change costs today, and how much review and rework it needs.

In the second week we change defaults together with the team, such as which model handles which kind of task, how much context an agent loads and where budgets sit. Then we measure the same kind of work again. A change stays only if the cost falls and review effort and rework do not rise.

• One team that already uses Claude Code, Cursor, GitHub Copilot or OpenAI Codex.
• Usage and billing exports from those tools, or read access to their admin consoles. We need no production data.
• A named contact in the team, and an hour with the engineering lead at the start and at the end.
• A confidentiality agreement, if you want one before we start.

We report by team, tool, model and type of task, and we do not rank individual engineers. Finding the savings rarely needs it. In German companies, measuring individual people can also require the works council's agreement before anything starts.

The review stands on its own, and you decide what follows once you have the result. If you continue, the services below cover the rollout to more teams, ongoing cost controls and training on your own code.