▶ Watch ↗AI Engineer World's Fair 202631:24
Conference affiliation: Member of Technical Staff · Anthropic · 2026
Gagan Bhat is a member of technical staff at Anthropic working on Claude Managed Agents and a core maintainer of the Model Hardware Standard. His technical writing and demonstrations explain how to give agents access to engineering systems without exposing credentials, losing their progress, or leaving developers unable to reconstruct their actions. His newer work concerns agents operating physical equipment, with safety evaluations being developed alongside early integrations.
Gagan’s early engineering work included FIRST Robotics teams, where he wrote Java and C++ code for autonomous robots, used motion sensors and navigation systems, and built dashboards to monitor machines during competition. At the University of Texas at Dallas, he researched temporal lidar data for semantic segmentation and autonomous robot exploration in unfamiliar environments. The university’s Intelligent Robotics and Vision Lab lists him among its undergraduate alumni.
His career history subsequently moved through consumer-device software and cloud gaming. During a 2021 Netflix internship, he worked on television applications in Device Ecosystem Engineering, crediting his mentor and manager for supporting the work. He also interned on cloud gaming at NVIDIA and on Apple’s iPhone Camera & Photos team. He worked on Netflix’s TV App SDK from May 2022 to August 2023, then on NVIDIA’s GeForce NOW before joining Anthropic in September 2025.
In his article with Isabella He, Gagan explains how the responsibilities of agent developers change across Anthropic’s interfaces. With the Messages API, an application calls the model, executes tools, returns their results, and repeats. The Claude Agent SDK packages the harness used by Claude Code. Managed Agents adds hosted orchestration and infrastructure for execution, credentials, and persistent sessions. Developers still define the task, tools, guardrails, and domain context; a managed harness does not supply the knowledge that distinguishes an incident investigator from a legal or sales agent.
Gagan makes that distinction concrete through incident-investigation examples. In his joint presentation with Isabella, he demonstrates an agent investigating a spike in checkout latency. It searches application logs, retrieves metrics and deployment history through MCP tools, and examines a code change to identify a likely root cause. Its network access is limited to the intended MCP server. The example connects the agent’s task to the evidence it needs and the systems it is permitted to reach. His Managed Agents workshop teaches developers to build a working incident investigator around agent, environment, and session definitions.
Two mechanisms are central to his explanations:
Gagan and Isabella also explain why harness workarounds need reevaluation after model upgrades. Context resets helped Sonnet 4.5 when it began finishing tasks prematurely near its context limit. Opus 4.5 no longer exhibited that behavior, so the resets became overhead. The example captures a specific engineering tradeoff: software added to compensate for one model’s limitation can hinder its successor.
Gagan has also appeared alongside Netflix engineers in a discussion of production AI infrastructure and evaluation, representing Anthropic’s Applied AI team. That appearance is distinct from his earlier Netflix employment.
Gagan’s public career history identifies him as a core maintainer of the Model Hardware Standard. In an original post on his public profile, he describes several months of work on the standard for AI agents operating physical devices, including equipment in scientific laboratories and advanced manufacturing.
He describes the research preview as a collaboration with partners to build safety evaluations ahead of a planned open-source release. That status matters: the work concerns developing and evaluating an interface for physical action, rather than establishing that agents can already operate such equipment safely across those settings.
▶ Watch ↗AI Engineer World's Fair 202631:24
Loading more talks and highlights…