Olive Shoot: An Agentic OS for Solopreneurs, Built in 6 Weeks

The brief

Solopreneurs run their businesses across 8 to 12 disconnected tools: notes, tasks, CRM, calendar, finance, content, comms. Every one of those is a tax. Olive Shoot is an opinionated, AI-native operating system that collapses the tax. One workspace, one schema, one set of agents.

What we shipped

A production-grade web application with:

• Real-time multi-tenant workspaces (Supabase RLS + Realtime)
• Type-safe end-to-end RPC (tRPC v11 with React Query)
• Server-driven UI with Zustand v5 for client state
• Embedded AI agents for triage, drafting, and research
• Stripe-backed billing with metered usage on AI calls
• Cron-driven background jobs via Trigger.dev

Six weeks from blank repo to public preview. Twelve weeks total to production.

The architecture choices that mattered

tRPC v11 over a REST API. Every minute spent debugging API contracts is a minute not spent shipping features. tRPC removed that minute entirely. Backend changes propagate to the frontend through the type system. We never wrote a single API integration test.

Supabase RLS as the auth boundary. Every query the frontend issues goes directly to Postgres through PostgREST. The only auth code we wrote lives in row-level security policies. No middleware layer to maintain. No service-account-with-service-key footgun.

Zustand v5 for ephemeral client state, RSC for everything else. We treated React Server Components as the default and pulled state down to the client only when interactivity demanded it. This kept the bundle tight and the data layer honest.

What this looked like in practice

Week 1-2: Schema design, auth, workspace primitives. Week 3-4: Core domain (notes, tasks, contacts), tRPC layer, base UI. Week 5-6: AI agent integration, embeddings, real-time presence. Week 7-12: Hardening, billing, cron jobs, observability, beta launch.

No standups. No estimates. No sprint planning. The roadmap was a pinned tab in Linear and a 30-minute weekly review.

What we learned

Shipping fast is a stack problem first, a discipline problem second. With the right primitives, the discipline becomes possible. Without them, no amount of discipline saves you.

Real-time is cheap if you start with it. Adding Supabase Realtime in week 2 was 90 minutes. Adding it in week 12 would have been a refactor.

AI features should be vertical, not horizontal. A "summarize" button on every screen is noise. A single workflow ("triage my inbox into next actions") is value.

Where it goes next

Public beta is open. The next push is multiplayer agents, where multiple specialized AI roles operate inside the same workspace coordinating through shared state. If you want a tour, book a 30-minute walkthrough.