LLMs as translation layers, not autonomous agents

Draft
  • engineering
  • llm

The reliable wins I've shipped with LLMs all share a shape: a known input on one side, a known consumer on the other, and the model doing the lossy translation in the middle.

Draft — full write-up coming. Outline below.

  • Agents promise autonomy and deliver debugging.
  • Translation layers: bounded input, bounded output, testable.
  • Examples from test infra and GTM tooling.
  • How to keep the blast radius small.