Next.js + Vercel AI SDK + Postgres/Supabase + Stripe is the fastest path from zero to paying customers for solo founders. One language (TypeScript) end-to-end, excellent deployment story (git push to Vercel), and the AI SDK handles streaming, tool calling, and multi-step agents out of the box.

Best when: your agent logic is moderate complexity (5-10 tools, single-agent architecture), you want to ship UI and backend together, and your team is JS-native. The Vercel AI SDK supports OpenAI, Anthropic, and Google models with unified tool-calling interfaces.

Typical architecture: Next.js app with API routes for agent endpoints, Supabase for auth + database + pgvector for RAG, Inngest or Trigger.dev for background jobs, Stripe for billing. Total monthly infra cost under $100 until you have hundreds of users.

FastAPI or Modal for agent workers, LangGraph for state machines, Next.js or React frontend. Best when your agent logic is complex (multi-agent orchestration, custom eval pipelines, heavy data processing) or when your team has ML/Python expertise.

LangGraph excels at stateful agent workflows — loops, branching, human-in-the-loop checkpoints, and persistence across steps. If your agent needs to pause, wait for human input, and resume, LangGraph handles this cleanly. FastAPI serves agent endpoints; Modal or Railway handles deployment and scaling.

Typical architecture: Python backend (FastAPI + LangGraph) deployed on Modal, Next.js frontend on Vercel, Postgres + pgvector for storage, Redis for task queues. More moving parts than the JS path, but better for complex agent logic that would be awkward in TypeScript.

Three infrastructure traps kill early-stage agent startups: custom vector databases, Kubernetes, and multi-region deployment. None of these matter until you have paying customers and proven product-market fit.

Use pgvector (built into Postgres/Supabase) for RAG — it handles millions of chunks and eliminates an entire service to manage. Use managed hosting (Vercel, Railway, Modal, Fly.io) instead of Kubernetes — you do not need container orchestration for 50 users. Deploy in one region until latency complaints arrive — multi-region adds complexity that delays shipping.

Your stack should let you deploy a prompt change in under 5 minutes. If changing a system prompt requires a Docker rebuild, Kubernetes rollout, and on-call engineer, your iteration speed is too slow. Speed of iteration beats architectural elegance at the MVP stage.