Sandhya Subramani works in applied AI, language models and agent systems, exploring how agents can select actions, acquire new software capabilities and control physical hardware. At AI Engineer World’s Fair 2026, she was a senior developer advocate for generative AI at AWS. Her Scout rover experiment makes agent orchestration tangible: conversational requests become calls to existing robot tools and policies, with the limits of physical execution visible alongside the successes.
Applied AI across business and research
Her background includes developing and deploying AI systems at Amazon, Warner Bros. and Fidelity Investments. Her speaker biography and session on self-evolving agents describe work in enterprise cybersecurity at Fidelity, machine-learning and AI research and development at Warner Bros. Discovery, and a lead applied AI scientist role at a stealth startup. These roles span business applications of machine learning and research into systems that can choose and execute actions.
In her account of an Amazon Renewed internship in 2020, Subramani describes developing an NLP-based tool to improve data quality, collaborating on product-selection monitoring and producing data-driven seller recommendations. The work placed machine learning within an operational problem: improving the information and recommendations teams used to make decisions.
Expanding an agent’s tools during execution
Her work on runtime capability expansion addresses a different limitation. An agent may encounter a task for which its existing tools are insufficient. Her self-evolving agent architecture explores generating a new tool during execution, validating it and incorporating it into the continuing workflow. Validation and safe incorporation matter because generating code alone does not establish that the new capability is suitable to use. The aim is to let an agent continue a task without first requiring a developer to redesign and redeploy the application.
Scout: connecting language to robot actions
In her Scout demonstration, Subramani extends tool use into robotics through Strands Agents, an open-source framework built by AWS. The central distinction is between choosing an action and carrying it out: the agent interprets the request and selects a robot function or learned policy; the controller or policy determines how the robot executes it. Asking a robot to pick up a red cube works only if it already has that physical capability. The agent can coordinate available capabilities through language, but it does not supply missing motors, controllers or learned skills.
Cooperating agents: Scout uses three cooperating agents. One continually assesses the environment and considers what to do next. Another handles conversation through web and Telegram interfaces and translates requests into actions. A third supports voice interaction; Subramani disabled it during the presentation because it would otherwise interpret her speech as instructions and interrupt. This division lets ongoing perception and user interaction operate together without treating every input as the same kind of task.
Execution layers: She explains the robotics system through four layers: the agent, a policy provider, an execution backend and the physical hardware. Instructions move toward the robot, while observations return toward the agent. Her proposed cloud-and-edge architecture combines cloud-based training and heavier computation with suitable functions and policies running at the edge for faster execution. The distinction matters in a physical system, where deciding what to do and delivering a timely movement command are separate engineering problems.
Training episodes: The rover also provides a way to collect training episodes. Subramani can guide its movement manually and record how it responds and reasons about requests, creating information that can inform later training and improvements. Interaction therefore serves both as a demonstration of existing behavior and as a means of studying how to improve it.
What the live test reveals
The live test preserves the difficulty of that work. Scout describes camera observations, responds through Telegram and performs a full turn, but it also takes time to respond, fails to carry out some requests and tumbles during movement. A request to do something complex prompts a speech-tool call without immediately producing the intended physical behavior. These moments show why fluent conversation cannot stand in for successful execution. Subramani presents agent coordination as an intermediate approach while VLA models improve; her longer-term ambition is for more capable models to reduce the need for task-specific robot training.
Sandhya Subramani shows how Strands Agents connects natural-language requests to existing robot capabilities—and how live execution exposes the gap between choosing an action and completing it.
An agent broadens how people request and combine robot capabilities. The agent selects what to do; an existing policy supplies how to do it, and missing physical capabilities remain missing.
The architecture separates agents, policy providers, execution backends, and hardware. Its proposed cloud-and-edge split combines cloud training with faster local execution, without establishing measured performance gains.
Natural-language fluency and completion messages are insufficient evidence of physical success. Scout’s complex-task response produces speech and a headlight change, while movement and recovery remain unreliable.
Manual navigation and conversational interactions can generate training episodes for later improvement. The demonstration establishes a way to collect data, rather than verified autonomous learning during use.
Camera-based conversation should retain the limits of the observation. Scout reports people counts, but its final answer also acknowledges that a low camera angle makes faces difficult to assess.
▶ Watch ↗