Jeremiah Lowin

Jeremiah Lowin is the founder and chief executive of Prefect and the creator of FastMCP, an open-source Python framework for connecting AI agents to tools and data. His career spans financial risk management, data orchestration, and agent infrastructure, with a consistent focus on making automated systems observable, dependable, and controllable.

Lowin studied economics and statistics at Harvard, earning a master’s degree in statistics while still an undergraduate. He joined King Street Capital in 2007 as the hedge fund developed its risk-management function, later founded the machine-learning consultancy Lowin Data Company, and subsequently became director of risk at another investment firm.

As maintaining analytical infrastructure overtook the analysis it supported, Lowin became an early Apache Airflow contributor and served on its project management committee. When Airflow proved insufficiently flexible for his data-science workflows, he built an internal alternative called Tin Man. Interest from other teams led him to found Prefect in 2018; its orchestration framework became open source the following year.

He extended that work into AI through Marvin, an agent and AI-application framework, before creating FastMCP. An early version entered the official Model Context Protocol Python SDK; FastMCP subsequently evolved into a separate, higher-level framework for building MCP servers and clients.

What agents need from software

Lowin treats an MCP server as a product interface for an imperfect user: an agent with limited working context, costly iteration, and incomplete knowledge of its tools. His AI Engineer session on MCP server design develops several practical principles:

• Agent-native product design: Design tools around what an agent can accomplish, using clear names, simple arguments, constrained choices, and errors that explain how to recover. Tool examples influence model behavior so strongly that even incidental details can shape subsequent outputs.

• Outcomes over operations: Encapsulate predictable sequences of API calls inside a single outcome-oriented tool. Conventional code should handle deterministic orchestration; agents are more useful when the necessary steps cannot be specified cleanly in advance.

• Context-aware tool curation: Every tool description and schema consumes finite context, so exposing an entire REST API can overwhelm an agent before it begins. Automatic conversion can bootstrap development, but production tools require selective exposure, concise documentation, and progressive disclosure where clients support it.

• Governance as infrastructure: Permissions, confirmation flows, observability, and deployment require actual infrastructure. FastMCP Cloud addresses hosted deployment, while Prefect Horizon manages how organizations distribute and govern agent-accessible tools and context.

Lowin supports interoperable, vendor-neutral standards through Prefect’s participation in the Agentic AI Foundation. His personal projects include copychat, which prepares code for model context, and claude-wt, which manages parallel Claude Code instances. In July 2026, he announced the FastMCP v4 beta, rebuilt on the newer MCP Python SDK.

That same month, Prefect acquired Dagster Labs, uniting two previously competing orchestration ecosystems. Lowin positioned Dagster around defining data outcomes, Prefect around reliable execution, and FastMCP around giving agents governed access to tools and information.