Three AI models working as one team. And the bill for every task.

01 · Constraint

Every AI reply paid again for the same context.

On every reply, Claude Code rereads the instructions and project files. It is like paying the entrance fee again for every question.

Heavy users could not see that cost, or tell whether the right model was doing each task.

02 · Decision

One conductor for three models, with rules that apply themselves.

1. AConstitution, the team handbookAbout 100 lines and 9 laws on how the agent works. Short on purpose, because it is reread every turn.
2. B6 hooks, the automatic inspectorsThey run on their own during the session. They log usage and warn when the cache is about to break or the rules get bloated.
3. CAudit, the statement for every sessionA single Python file that reads your session logs and computes cost per turn, cache and split by model.

The golden rule: simple tasks go to the lightest, fastest model, hard tasks go to the most capable. Each type of agent has its model.

03 · Accepted cost

What was left out, on purpose.

• Hooks warn, but never block. The person still decides.
• Installation has three phases. The last one only runs after you confirm you have read the whole constitution.
• It is not for everyone. For people using it less than an hour a day, keeping it running costs more than it returns.

04 · Result

90 days of real use, measured reply by reply.

987

sessions

201,910

model replies

71.76%

of the work done by sub-agents

91.63%

cache hit rate

One operator, 90-day window, Jan 29 to Apr 28, 2026. Source: cost-audit.py output and EVIDENCE.json in the repository. Your numbers will differ.