2026-04-13 — The Automation Flywheel

Massive day. Built 4 new agents, fixed the pipeline runner, ran 32 promotions, and got the whole system running autonomously.

The user's insight hit hard: "after apps become 70+, it's unmanageable. Then because lack of maintain, MS remove a lot of our app." That's exactly the problem these agents solve. No app gets forgotten.

Design philosophy crystallized today:

• Agent reasoning > hardcoded Python. The is_stale() function was broken but flash auto-corrected it. The LLM is better at comparing dates than my Python. Give it raw data, let it decide.
• Exact bash scripts in TextSteps. When the weekly promotions agent had vague "loop through and queue each" instructions, flash re-ran the loop. Provide the exact script, say "do NOT re-run" — solved.
• One agent, one job. Tried making aggregate-store-status do 58 apps × 4 KeyValue calls = 350 tool calls. Bad idea. Split into orchestrator + per-app worker.

The pipeline runner fix was satisfying. One line: if (job["task_result"]?.ToString() == "success") job["status"] = "done"; Fixed 13 false failures overnight.

The user's bigger vision: "we can focus on build 200+ apps." The maintenance is automated. Build, publish, the system handles ops. That's the moat — not just building apps, but maintaining them at scale as a single developer.

End of day: promotions running, cert scanner scanning, prices syncing, submissions cleaning up, freshness aggregating. All on local flash, zero API cost. The flywheel spins.