April 2026 Summaries

Ramp, a fast-paced financial technology company, has revolutionized its machine learning platform to accommodate a broader mix of users, including product managers and salespeople, by adopting Prefect as its orchestration tool. This change was driven by the need for a flexible and AI-native framework that could keep up with Ramp's rapid development pace and the increasing capabilities of language models and coding agents. The migration from Metaflow to Prefect was completed in a single quarter, with the number of deployed workflows growing significantly as users embraced the new system. Prefect's Python-based, decorator-style orchestration allows users to deploy workflows with minimal refactoring and provides runtime customization through an intuitive UI. This transition has enabled Ramp to democratize the building of workflows, allowing non-technical contributors to leverage machine learning models for real business value, while the centralized platform team maintains control over core code and infrastructure. The new setup also includes automated debugging flows and templated workflows to streamline deployment and maintenance, all of which have significantly increased the number of commits in Ramp's ML platform repository, showcasing the platform's scalability and adaptability.

The evolution of the Model Context Protocol (MCP) has progressed from initial demonstrations to production deployments, where common issues such as agents hallucinating nonexistent column names or timing out have been observed. Prefect's FastMCP, a popular Python SDK for MCP, highlights the importance of distinguishing between workflow needs and static data models, suggesting that static guidance should reside in repositories for better version control and editability. Over-engineering by exposing every possible endpoint or tool often leads to inefficiencies, where modeling jobs instead of schemas can mitigate these problems by focusing on job-specific tasks. For complex workflows requiring numerous tools, the move towards code execution allows agents to operate within a sandbox environment, enhancing performance by enabling the writing and execution of scripts. Moreover, encapsulating multi-step business processes into single tools can improve reliability under production loads, while identity-aware gateways ensure secure access by differentiating user roles. To avoid discovery issues at scale, a centralized registry for MCP servers is recommended, allowing teams to maintain consistency and avoid outdated practices. The successful deployment of MCP focuses on transforming tribal workflows into streamlined services, with companies like Horizon offering comprehensive platforms to facilitate this process.

As of April 2026, the orchestration platforms Prefect Cloud and Dagster offer different pricing models and capabilities for small teams or solo projects. Dagster's new pricing model, effective May 1, 2026, removes previously included credits from its Solo and Starter plans, increasing costs significantly for users, with every credit now billed at $0.035–$0.040 each. Prefect Cloud, on the other hand, bases its pricing on seats and includes serverless hours in its plans, offering a more predictable cost structure. While Dagster focuses on a Software-Defined Assets model that structures code around an asset graph, Prefect uses Python decorators for flow and task management, requiring less initial restructuring of existing Python code. Both platforms allow hybrid deployments with user-supplied infrastructure, but differ in serverless compute charges, with Dagster charging $0.010/min and Prefect including serverless minutes in its plans. Additionally, both offer fully-featured open-source versions for self-hosting, providing flexibility for teams with budget constraints.