AI agents rival industrial policies in zero-shot navigation with minimal interface

A new paper by Jian Zhou and colleagues explores agentic embodied control—where a general-purpose AI agent maintains the decision-making loop (perceive, act, verify, self-correct) rather than relying on task-specific workflows or trained policies. Using zero-shot navigation as a testbed, they evaluated three software-engineering agent harnesses with only a monocular RGB camera and discrete actions. Under this minimal interface, the opus-5 model achieved 70.7% success with default settings, while fable-5 reached 78% at maximum effort. A hybrid agent that exposed a trained waypoint tool as an optional capability achieved 76.7% success at default effort, using half the environment steps and less than a quarter of the wall time compared to the maximum-effort primitive-only run.

Model choice dominated success rates, while harness choices were descriptive—i.e., they affected behavior but not core capability. Notably, forcing a waypoint interface helped weaker models but hindered stronger ones. However, performance fell sharply on longer-horizon tasks, and latency plus context growth remain barriers. The results show that general-purpose, zero-shot agents can now compete with specialized industrial-scale policies, and that complementary improvements in models, harnesses, and interfaces can advance autonomous embodied agents further.