AI-CEO Framework: only 4 of 11 departments actually run

The author's setup is unusually literal: .claude/agents/ holds 17 agent definitions and .company/departments/ holds 11 department directories, each with a STATE.md file as its living state. What broke the correspondence between the two was not enthusiasm but a cost structure. When he measured the real cost of unattended claude -p runs in September, he found a single call cost $0.27–0.47, and cache creation accounted for about 41,000 tokens. Only 5,800 of those were his own files; the rest was Claude Code's fixed overhead, meaning prompt trimming saves little and only the number of calls is genuinely reducible. He then routed every unattended run through one wrapper, claude-run.sh, that captures real cost as JSON, caps each call at --max-budget-usd 3, and tags records so that monthly spend per department becomes visible. Scheduling follows the wrapper: a monthly auto-agent-health-check.sh simply looks at STATE.md modification dates and flags any department untouched for more than 30 days. As of October 7, publishing, marketing, sales and dev had recent updates and their own jobs, while consulting, finance, legal, tax, cs and hr had neither. The author is explicit that stopping those departments was the right call, and he now advises counting revenue-generating business lines first — publishing, contract development and SaaS in his case — and creating departments on a one-to-one basis with them, leaving accounting and legal to be called as skills when needed. His other warnings are operational: a CLAUDE.md that grew past 1,000 lines began producing instructions opposite to what was written, so orchestrator instructions are now kept under 200 lines with department-specific detail split into each agent definition; giving a department agent Write permission on a schedule can lead it to rewrite files instead of producing a report, so recurring report jobs are pinned to --allowedTools "Read,Bash" and only writing departments go through an approval queue. Verification, he argues, is a matter of logs — the last run time and the number of cost records — because if both have stopped, the department effectively does not exist.