Harness Engineering: How to Keep Agents on Track

Most agentic coding discussions obsess over the model. This talk is about the box around it: the harness. Tools, permissions, context, observability and feedback loops are what turn an LLM from a text generator into something that can work inside a real codebase. I walk through how coding agents actually work, why code is an unusually good fit for feedback-driven systems, and why platform engineering is the right analogy for scaling agents across teams. You'll leave with a practical checklist: give agents senses, narrow the blast radius, treat context as scarce, and wrap nondeterministic behaviour in deterministic checks. The harness matters more than the model.