Automating the Design of Embodied Agent Architectures

This paper studies whether Agent Architecture Search (AAS) can transfer from text agents to perceptual embodied agents. It introduces AgentCanvas, a typed-graph runtime for editable node-and-wire executors with simulator-aware execution and episode logs, and KDLoop, a coding-agent search process involving proposal, critique, experimentation, distillation, and stall-triggered reflection. Three AAS variants are evaluated across four embodied executors covering vision-language navigation, embodied question answering, and language-conditioned manipulation. The reported 3x4 matrix shows deployable, directional success-rate gains, while also revealing rollout noise, local edit basins, incomplete credit assignment, and a leak-bearing high-scoring candidate.