Bow: case study notes

Moved out of the README to keep it short. Same rule applies here: receipts over hype.

What this demonstrates

Transferable signal for an AI-builder reader:

• Systems design under real constraints. Auth, cost, macOS TCC, concurrency, reliability, all solved against the wall, not on a whiteboard.
• An AI-native SDLC. Spec → plan → subagent build → independent adversarial QC → security review → tests → reversible production cutover.
• An original orchestration doctrine (Fleet Mode) that generalizes how to run agent fleets, directly transferable to anyone building agentic products.
• Build-vs-buy and cost judgment. Collapsed a multi-tool, multi-provider stack to one AI at about $0 marginal cost over an existing plan.
• Honest engineering communication. A scorecard with explicit misses and a real limitations section.
• Shipped and operated. It runs daily and absorbed a live job via an atomic, reversible cutover. A redacted tail of the live daemon log, with real ISO timestamps showing a routine firing and the watchdog confirming the heartbeat, is in assets/daemon-log-excerpt.txt.

A note on the artifacts in assets/: the pytest output and the daemon log excerpt are raw tool output (only secrets, ids, and paths redacted). The FLEET MODE and QC receipts are hand-formatted summaries of real runs, clearly stylized, not raw console capture. The two kinds are kept visibly distinct on purpose; conflating a formatted summary with raw evidence is exactly the kind of thing this project is trying not to do. Raw: receipt-tests, daemon-log-excerpt. Formatted summaries: receipt-fleet-mode, receipt-qc-catch.

Cost

About $0/mo over an existing flat-rate plan, vs. metered API pricing that is nonzero and recurring forever for an equivalent do-everything agent.

The build-vs-buy reasoning is the whole architecture in one line: a do-everything agent built on a metered API key costs real money every month, forever, on top of a flat-rate plan I already pay for. The exact metered figure depends on traffic, so I won't dress a guess as analysis. The load-bearing point is the shape: a recurring monthly bill that scales with use, versus about $0 marginal. As one worked anchor, a do-everything assistant running all day at even ~1M Opus-4.8 tokens/day of mixed input/output lands in the low tens of dollars a month at current metered rates and climbs from there with heavier build traffic. Recurring, forever.

By being a first-party Claude Code user rather than a third-party API caller, Bow puts every chat turn and every build under the plan I already have, and inherits skills, memory, MCP servers, and tools for free instead of rebuilding two-thirds of the old system. That is why the marginal cost is about $0 and why two-thirds of the old feature list was never rebuilt: it wasn't needed.

Closing

Full private repo and the test suite available on request.

I architected this system, made the central bet (be the tool, don't call the API), and then directed and gated fleets of Claude (Opus 4.8) subagents to implement and adversarially review it, under the Fleet Mode doctrine, which runs as a live skill so the orchestration is operational, not an essay. The gates caught real bugs every single time; the soft-lock catch in the README's section 5 is the proof that orchestration-under-gates earns its keep.