Sarmad Qadri is the co-founder and former chief executive of LastMile AI, creator of the open-source mcp-agent framework, and an AI infrastructure builder at Meta Superintelligence Labs. He designs AI systems whose agents can share a common interface, coordinate specialized tasks, recover from failures, and continue working after users close their applications.
From developer tools to production AI
Qadri studied computer science at the University of Waterloo, earning a Bachelor of Mathematics. At Microsoft, he worked on compilers, build systems, and developer tools, including the Language Server Protocol and language servers between 2016 and 2018. He subsequently joined Meta, helped build Facebook’s Jupyter notebook platform, and became an engineering leader working on machine-learning experimentation and model-management infrastructure.
In 2023, Qadri founded LastMile AI with fellow former Meta engineers Andrew Hoh and Suyog Sonwalkar. As chief executive, he helped raise a $10 million seed round led by Gradient to develop production-oriented AI tools for software engineers. The company’s open-source AIConfig project put prompts and model settings into version-controlled configuration, making AI application behavior easier to inspect, test, and revise.
Making agents interoperable and dependable
Agents as MCP servers. Qadri applies the logic of the Language Server Protocol to the Model Context Protocol: standardized interfaces should allow independent clients, tools, data sources, and agents to communicate without bespoke integrations. His mcp-agent framework exposes agents themselves as MCP-compatible services, allowing clients to invoke specialized agents that can call other agents and tools through the same protocol.
Durable asynchronous agent workflows. Qadri models agents as persistent workflows that can pause for human feedback, retry failures, resume later, and respond to scheduled events. Temporal-backed execution separates an agent’s work from the application that initiated it; an orchestrator can assign retrieval, proofreading, fact-checking, and writing to specialized agents without requiring an uninterrupted chat session.
Deterministic plan verification. While building a deep-research orchestration system, Qadri found that sophisticated adaptive workflows could still fail practical tasks. His Deep Orchestrator instead checks that proposed agents, MCP servers, and task dependencies actually exist, propagates context selectively, and replans when necessary.
Qadri and the LastMile team subsequently joined Meta’s Products & Applied Research organization within Meta Superintelligence Labs, where he is developing long-running, interconnected agents for executives and knowledge workers.
