Gagan Bhat is a member of Anthropic’s Applied AI technical staff working on Claude Managed Agents, infrastructure for building secure, reliable AI agents. His work addresses the operational challenges separating useful prototypes from production systems: execution environments, credentials, observability, persistent sessions, and specialized context.
Bhat studied at the University of Texas at Dallas, where he worked with its Intelligent Robotics and Vision Lab, and held engineering roles at NVIDIA and Netflix. At Netflix, he worked on television applications in Device Ecosystem Engineering, an experience he described in a public account of his internship. His subsequent work at Anthropic combines product development, model evaluation, and collaboration with enterprise customers, including production agent development at Netflix.
Building agents that survive production
- Production-grade agent infrastructure: Bhat distinguishes the Messages API, the Claude Agent SDK’s packaged agent loop, and Claude Managed Agents’ hosted execution, credential management, and observability. Developers retain responsibility for the product, task, tools, and domain knowledge.
- Isolated execution and credential protection: Sandboxed environments restrict network access and tool execution, while secure vaults decrypt credentials only when needed, keeping authentication secrets away from the model.
- Durable session logs: Persistent records of requests, responses, tool calls, and results support debugging, recovery, and personalized agent memory. Bhat calls the periodic conversion of session histories into improved memory dreaming, an experimental direction that could eventually incorporate organizational knowledge.
He applies these principles to an incident-investigation agent that examines application logs, deployment history, and operational metrics. His writing with Isabella He frames adaptable infrastructure as essential when rapidly improving models change what agents can accomplish.
