AI Agents Need Harness Engineering, Not More Hype

Most teams still talk about AI agents like the hard part is getting the model to call a tool. It isn't. The hard part starts right after the demo works once: coordinating state, handling failures, deciding what gets retried, keeping context coherent ...