YC: harness design changes agent behavior beyond model weights

Y Combinator’s State of Harnesses session examines the system around an AI model: persistent context, tools, sandboxes, agent-to-agent messaging, budgets, and verification. Its practical argument is that these choices can materially change how the same model performs on multi-step work. Why it matters: Choosing a stronger model will not fix an agent that loses its working state, has the wrong tools, or can act without limits. The operating layer deserves the same design and testing attention as the model choice.

Try this: For one multi-step agent workflow that fails intermittently, trace what the agent receives on two consecutive turns. Check whether prior tool results, decisions, and action limits are still present before changing models.