Keeping the Thread: Human Understanding and Responsibility While AI Agents Act

As AI systems move from providing advice to taking action, human oversight must remain usable while those systems operate. This independent working paper introduces "keeping the thread": maintaining continuity of human understanding and responsibility as AI agents act, circumstances change, and people move in and out of direct involvement. Drawing on examples from aviation, healthcare, and mission control, the paper identifies six conditions for meaningful human oversight and outlines design principles for preserving the ability to understand, intervene, and verify outcomes. It also considers people affected by AI decisions who may never interact directly with the systems involved. The paper builds on independent design research and contributions to ITU-T FG-TIDA during 2026.