Jens Fischer

Build vs Buy

When to call an API, when to use a framework, when to host your own model.

The default should be: call the smallest abstraction that gets the job done. Most "AI platforms" you're considering buying you don't need.

The ladder

  1. Just the API. Anthropic or OpenAI SDK. A while loop. This covers 80% of real applications. Start here.
  2. A thin framework. Vercel AI SDK, Mastra, Pydantic AI. Useful when you want streaming, tool-call plumbing, or typed schemas without writing them yourself.
  3. An orchestration framework. LangGraph, Inngest, Temporal. When you need durable state, human-in-the-loop, or branching workflows.
  4. A managed agent platform. LangSmith, AgentOps, Vellum. When ops, evals, and observability matter more than control over the loop.
  5. Self-hosted open model. Llama, Qwen, Mistral on vLLM or Ollama. When latency, privacy, or unit cost at scale forces the question.
  6. Fine-tuning. Almost always premature. Try prompting and retrieval first.

Decision framework

  • If the problem is prompt engineering, no framework helps. Skip to evals.
  • If the problem is orchestration (long workflows, retries, durable state), buy or use a framework.
  • If the problem is observability, buy. Building this from scratch is a tarpit.
  • If the problem is regulatory (data residency, no third-party processors), self-host.
  • If the problem is cost at scale, run the math. Hosted models are cheaper than commonly assumed under ~100M tokens/month.

What to avoid buying too early

  • Vector DB before you have a corpus that needs one. Postgres + pgvector handles ~1M rows.
  • "AI gateway" products before you have a routing problem.
  • Fine-tuning services before you've exhausted prompting and RAG.