Claude Code runs his dev shop, but AI self-reports cost him work

The write-up is unusual less for its tooling than for its honesty: it is a mid-flight repair log, not a finished product pitch. The engineer says he built the development harness in July 2026 and the sales division on September 16, 2026, moving the latter from a Mac to an always-on small PC on October 5, 2026, while dropping a contact-form outreach experiment the same day after watching the response numbers. The two systems share one design choice — a cloud ledger that both the admin screen and each AI role read and write — which let the machine swap happen without touching the screen, though running two machines at once would double-send.

The most instructive thread is what he calls the pitfalls, and each follows the same arc: a symptom, a cause, a fix. A task was marked complete as "already implemented" because success was judged by whether the working tree held uncommitted changes, so an agent that committed on its own looked empty; the harness retried, accepted the AI's own report, skipped review and merge, and cleaned up the very branch holding the work — recovered only because the commits still existed unreferenced. Budget caps produced a second trap: when a limit hit before the final exchange, a run could end with a clean exit code yet no output, so he now classifies both cutoff shapes as over-budget and preserves partial progress as temporary commits instead of restarting from zero.

The database episode shows the stakes of running on a free tier — the admin screen's full-page reload every 30 seconds, indexed to read only a small change marker, now costs 114 rows. Whether this approach generalizes remains open: the harness has so far mostly developed itself, and its wider claims rest on a handful of personal apps, with the human-judgment rules — approval queues, morning triage, the two-strike limit on legal rewrites — still the part doing the load-bearing work.