drop&grow — Case Study by Bankole David

the problem

Most idea apps are graveyards: you capture a thought and it sits still forever. And now that people and their agents work together, products still force agents to act through brittle browser automation instead of first-class tooling.

what i built

Built drop&grow: an agent-native idea space. You drop an idea any way you like, six specialized agents: Nova (research), Palette (design), Quill (content), Circuit (tech), Apex (strategy), Ledger (budget), build on each other into one coherent plan. The idea keeps living with branches, automatic connections, and a real-data health score, and every action a human takes, an agent can take through 26 exposed WebMCP tools.

• →Drop ideas by text, real-time voice, or image
• →Six-agent cascade: research, design, content, tech, strategy, budget
• →Branching evolution history, not arguments
• →Health scores grounded in real data, with a finish line
• →26 WebMCP tools so agents operate like any user
• →Handle-based sessions with secret-phrase protection

tools, and why

Convex for reactive, typed end-to-end state with serverless actions, Cron, and blob storage. React + Vite for a fast SPA. Groq with a Cloudflare Workers AI fallback for a speed-first agent cascade that never fails. Vitest + convex-test for 69 unit, integration, and e2e tests.

the hard part

Keeping six chained LLM passes to ~100–200ms without a budget. Solved with a provider cascade (Groq for speed, Workers AI on fallback, deterministic engine as the floor) so the swarm responds fast and never silently dies.