Summary
Sandhya Subramani, an AWS engineer, presented a live demonstration at the AI Engineer conference showing how to give a physical rover named Scout an agentic brain using AWS Strands, an open-source agent orchestration framework. Rather than pre-programming specific behaviors, Scout runs three simultaneous Strands agents: a thinker agent that continuously perceives its environment and decides what to do next, a communication agent connected to both a Telegram bot and a web app for natural language instructions, and a voice agent for real-time spoken commands. All three coordinate in parallel, with the system running via a 4G-connected Raspberry Pi demonstrated live on stage.
The architecture requires just five lines of code to connect the Strands agent layer to the robot’s preset hardware functions (called policies), allowing the LLM to decide which policy to invoke at runtime based on natural language input. Under the hood, VLA (vision-language-action) models trained in simulation serve as the policy layer, with Strands handling orchestration between cloud-based training and edge-based inference for low-latency physical execution — a hybrid model that enables both large-scale training and fast on-device response.
Subramani positions this as a generational shift in robotics: moving from rigid pre-programmed automation toward robots that understand open-ended instructions and adapt to novel situations without retraining. The demo showed Scout responding to commands it had never been explicitly trained for, including answering questions about what it observed through its camera, illustrating how the agentic layer extends a robot’s effective capability envelope far beyond its programmed policy set.
📺 Source: AI Engineer · Published August 29, 2026
🏷️ Format: Keynote Launch
