Mahesh Murag

Mahesh Murag is a member of Anthropic’s technical staff and a co-creator of Agent Skills, a system for giving AI agents reusable, specialized expertise through instructions, scripts, and reference materials. His work addresses three practical requirements for capable agents: connecting to external systems, applying domain-specific knowledge, and retaining useful experience.

At the University of California, Berkeley, Murag contributed to Flow’s research on traffic systems and autonomous vehicles. He subsequently worked at Scale AI on annotated machine-learning data, then joined Tecton, where he became a product manager overseeing compute-platform teams.

At Tecton, he worked on Rift, a Python-native managed compute engine for developing and deploying machine-learning features across batch, streaming, and real-time workloads. The platform addressed a familiar production tension: giving data scientists flexible Python workflows while maintaining reliable orchestration and low-latency serving.

At Anthropic, Murag applied that infrastructure perspective to the Model Context Protocol, then co-created Agent Skills with Barry Zhang and other colleagues. His more recent work explores memory and dreaming for self-learning agents.

Contributions and technical ideas

• Model Context Protocol as shared infrastructure. Murag frames MCP as a standardized interface between AI applications and external tools, data, and services. He distinguishes model-controlled tools, application-controlled resources, and user-controlled prompts, preserving separate responsibilities across the agent, application, and user. His architectural concerns include authorization, server discovery, composability, evaluation, and the security implications of third-party integrations.

• Agent Skills as portable expertise. Murag co-authored Anthropic’s engineering guide to Agent Skills with Barry Zhang and Keith Lazuka. Skills package procedural knowledge, scripts, and supporting files around a SKILL.md document, making specialized capabilities reusable, shareable, and accessible to domain experts outside software engineering. The format is also documented as an open Agent Skills standard.

• Progressive disclosure. Agents initially see concise skill metadata and retrieve detailed instructions or supporting resources only when needed, preventing large capability libraries from overwhelming the context window. Murag applies similar thinking to searchable and hierarchically organized tools.

• One runtime, many capabilities. A general-purpose agent equipped with code execution and a filesystem can combine MCP connectivity with task-specific skills instead of requiring separate scaffolding for every workflow. His subsequent work on memory and dreaming extends that architecture toward agents that retain, verify, and improve useful knowledge between sessions.