Building AI harnesses is just software architecture with worse documentation and higher stakes. The same mistakes that kill distributed systems — tight coupling, unclear failure modes, no observability, no graceful degradation — kill agentic pipelines. Except in a pipeline, the failure mode is a confident hallucination shipped to a customer instead of an exception that logs and stops.
Claude Opus 4.7 didn't introduce the problem. It made the architectural gaps visible at a level of capability that couldn't be ignored. When the model can execute complex multi-step tasks, the harness design stops being an afterthought and starts being the thing that either works or doesn't.
Most harness engineers building today have never shipped a system that needed to survive production load for two years. They know prompt engineering. They don't know capacity planning, circuit breakers, or what happens when the LLM endpoint degrades at 3am on a Tuesday.
Software architecture isn't a sub-skill of harness engineering. It's the main skill. Everything else is configuration.
Read the full piece on Medium →