• The churn loop does real forensics: for each cancellation in the last 24 hours it pulls Stripe payment and user data, watches the PostHog session replay, and checks Supabase usage to classify wrong customer fit, no 'wow' moment, or bug.
• Session replays are the highest-yield input in the whole system. Rage clicks and visible confusion show you where users struggle in a way that support tickets, which only capture people who bothered to write in, never will.
• Cloud routines beat self-hosting for a non-obvious reason: MCP connectors configured in the Claude desktop app are usable by both the local CLI and the cloud runs, so scheduled agents inherit integrations you already set up.
• Sequencing advice: build only the work-discovery half first, so routines file high-quality issues into the tracker you already use, and keep implementing however you do today. The leverage arrives before the loop is closed.

• MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.

If you're trying to run AI agents on real software work without babysitting them constantly, the insight here is structural: use Linear (or any issue tracker you already use) as a hard seam between work discovery and work execution, so you can extend either side independently without rebuilding everything. It's a practical blueprint you can actually replicate in an afternoon using your existing Claude subscription.

• postWhat Is a Software Factory?Chamath Palihapitiya
• articleAgent Native DevelopmentFactory
• clipThe software factory loop: signal to deployAI Engineer

Checking sign-in…

Loading comments…