The Current State of Context Engineering: From Long Context to Agent-Driven Harness

In a talk at AGNTCon, the author reviewed the evolution of context engineering, pointing out that long context brings four failure modes—context pollution, distraction, confusion, and conflict—and offered six fixes: context augmentation, tool loading, isolation, trimming, offloading, and summarization. As models shift toward autonomously driven harnesses, two new problems have emerged: "adversarial weights," where models overfit to their training environments, and "context inertia," where models become overly fixated on a single goal. Research from Anthropic, Microsoft Research, and UC Berkeley all addresses these by periodically clearing context while retaining environment artifacts. The author predicts that the next challenge will be memory accumulation and conflict in personal agents, which will require strategic forgetting.