Office Hours VIII (platform label 'XIII'): App-Store Economics, Vetting a Co-Founder Without Your Code, Android Accessibility Agents, and OpenClaw vs Hermes

Cross-platform first: React Native or Flutter, or a vibe-coding tool tuned for both stores (heard as 'Rork'; Replicate and Emergent named in passing). Then the money: Apple takes about 30% of in-app purchases and Google 15-30%. The workaround is Audible's - sell subscriptions on a web portal outside the app, and have the mobile app simply reflect the same database. A flat paid-app licence fee is a different model from in-app purchases and worth keeping distinct.

Share a pitch or a one-page brief rather than the code; or build a deliberately buggy dummy project and hand that over as a test assignment. You learn how they work, what they find and how they communicate, and your real IP never leaves the building. Shaquille's adjacent question - what to stress-test before production - gets the baseline four: latency, concurrency, load and ping; the rest depends on the product.

Cody's agent clicks on the user's behalf through Android's accessibility service, referencing a repo heard as 'agent Tars'. Akhil's reference architectures are Google Assistant and Apple's Shortcuts: a trigger phrase, a stored sequence of steps. Concretely: a JSON or markdown memory file mapping phrases to accessibility action sequences; a small fast model (Gemini Flash) for the which-tool decision, because two-minute on-device inference is unusable; and acceptance that OEM fragmentation is unavoidable - Redmi blocks accessibility, Samsung permits it - so ship two apps: a 'pro' setup app that handles permissions and a simple end-user app. Store approval: Apple straightforward with disclosure, Google slower and stricter.

The standard pattern is OpenClaw: expose the local model through its web interface and connect WhatsApp, Telegram and email as front ends. To reach localhost from the internet you tunnel with NGINX - which 'is not super secure, more like opening a pipe'. Akhil's own preference is Hermes, whose self-learning loop-engineering harness he calls 'much better' for the same job; Stefan's earlier question about wiring Hermes into an existing marketing workflow is pointed at the Circle recordings on Hermes and loop engineering.

Two modes. Step-by-step prompting is directing a junior intern; loop engineering is briefing a senior architect - you give well-defined goals, features, constraints and the tech stack, and let a high-capability model (Fable was named) design the path. Control comes from the brief, not from the steps. In Claude Code or Cursor you can assign different models to different sub-agents - a 'senior lead' and a 'junior intern' - and mix both styles in one system.