OpenAI's recent write-up describes building a 1-million-line application using AI agents with no manually typed code, relying on a "harness" of tooling and practices to maintain quality. The harness includes context engineering (knowledge bases, observability data), architectural constraints (custom linters, structural tests),

6m read timeFrom martinfowler.com
Post cover image
Table of contents
Harnesses - the future service templates?The runtime has to be constrained for more AI autonomy?A convergence on a limited number of tech stacks and topologies?Two future worlds: Pre-AI vs post-AI application maintenance?What’s your harness today?Final thoughts

Sort: