Automated tools can now propose, test, and refine the internal designs of AI agents that navigate physical environments - with mixed but meaningful results.
Researchers introduced two systems: AgentCanvas, a graph-based runtime that represents embodied agent architectures as editable node-and-wire programs, and KDLoop, a search procedure that cycles through proposal, critique, experiment, and distillation phases. Together, they extend Agent Architecture Search - previously studied mainly for text-based tasks - into agents that must perceive and act in simulated physical worlds. The team tested three search variants across four agent types covering vision-language navigation, embodied question answering, and language-conditioned manipulation. Results showed real success-rate gains in several configurations, though one high-scoring candidate was disqualified after the researchers found it had access to leaked information.
The significance here is methodological. Human researchers currently design these agent architectures by hand - deciding how memory is structured, how visual input is processed, how planning connects to action. Automating that design loop could compress development cycles and surface configurations humans would not think to try. The catch is that physical simulators introduce noise that text benchmarks don't, making it harder for the search process to distinguish a genuinely better architecture from a lucky rollout.
The paper is candid about what does not yet work: search can get stuck in local optima, and assigning credit to specific architectural choices across long episodes remains an open problem. That honesty is worth noting - most architecture-search papers bury their failure modes in appendices.