Embodied Agentic Harnesses for Physical Intelligence A Survey

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Embodied agents depend on more than the capabilities of a language, vision-language, or action model. They require a runtime that grounds observations, coordinates executable capabilities, monitors outcomes, and manages intervention when physical execution diverges from a plan. This survey examines embodied agentic harnesses as this runtime control plane. We distinguish four logical responsibilities—reasoning, runtime governance, skill execution, and hardware interfacing—and separate essential execution services from optional prediction, persistent memory, and adaptation mechanisms. A contract-based formulation connects skill preconditions, resource ownership, interruption, postconditions, and trace collection without assuming that every system implements an identical stack. We organize the literature by mechanisms rather than publication period: grounding and composition, state and memory, scheduling and coordination, verification and recovery, and adaptation. Complementary descriptors record the execution interface, intervention timescale, adaptation target, and evidence setting. The synthesis connects LLM and agent foundations, VLA policies, world models, and predictive-action models to affordance-based planning, programmatic control, memory-guided orchestration, agent operating systems, and evolving runtime critics. It distinguishes operational harnesses from infrastructure for robot learning and deployment. We further distinguish semantic task verification, invocation authority, and physical safety assurance, which require different evidence. The resulting synthesis identifies execution contracts, compositional evaluation, bounded adaptation, and explicit human authority as priorities for reliable long-horizon physical intelligence.