Nitya Narasimhan is a senior AI advocate on Microsoft’s Core AI Developer Relations team, helping developers build applications with Microsoft Foundry and GitHub Copilot. Her work concentrates on agent observability: making autonomous systems inspectable, measurable, and safer as their models, tools, and requirements change.
Narasimhan studied electronics and communications engineering at Bangalore University, worked at Infosys, and earned a master’s degree and doctorate in computer engineering at the University of California, Santa Barbara. Her doctoral research addressed middleware for reliable distributed systems. At Motorola Research, she spent more than a decade developing mobile and ubiquitous-computing technologies, became a Distinguished Member of Technical Staff, and accumulated 14 granted patents.
She subsequently worked in consulting, taught mobile development, led engineering at Cheerity, and ran Digital HV. In the New York region, she helped organize Google Developer Group NYC and AI Developers & Entrepreneurs before moving into Microsoft developer advocacy. Her focus progressed from mobile, web, and cloud development toward generative AI and responsible agent deployment.
Narasimhan describes herself as developer zero: testing emerging products early, uncovering friction, and translating developers’ needs into useful feedback for product teams. Her public writing on developer empathy emphasizes practical examples, documentation, and collaboration between engineering and developer-education teams.
What she brings to agent engineering
- Trace-linked agent evaluation: Connect evaluation failures to specific execution steps, including intent resolution, tool calls, grounding, and task completion. This makes regressions caused by model, prompt, or tool changes faster to diagnose. Her agent-observability workshop with Amy Boyd demonstrates custom trace attributes, OpenTelemetry, Application Insights, and Azure Monitor.
- Observable multi-agent workflows: Break complicated tasks into specialist agents whose handoffs, performance, and costs can be inspected individually. A travel-planning system can delegate flights, hotels, and car rentals while exposing precisely where an orchestration failed.
- Human-guided agent optimization: Use coding agents to generate evaluation datasets, establish baselines, improve instructions, compare versions, and identify regressions. Human oversight remains essential when improving one metric damages another or when an earlier agent version performs better.
- Adversarial testing for agent actions: Test whether manipulated prompts bypass safeguards, expose sensitive information, or trigger prohibited actions. Agents with external tools require scrutiny of what they can do, not merely what they say.
Narasimhan also creates illustrated technical explanations through SketchTheDocs and documents hands-on experimentation in her Nitya Learns AI Agents repository. Both projects extend her approach to developer education: make complex systems understandable through visual explanation, open experimentation, and reusable examples.
