Microsoft used the opening of Build 2026 to launch Rayfin, an open source SDK and command line that lets developers, and increasingly the coding agents working alongside them, describe an entire application backend in code and push it directly into Microsoft Fabric. The framing from Shireesh Thota, corporate vice president for databases at Microsoft, is that Rayfin turns backend development into a code first workflow covering databases, business logic, APIs, identity and access policies in a single deployable artifact. The repository is live at github.com/microsoft/rayfin and Microsoft is bundling a sixty day Fabric trial plus a Replit integration to lower the on ramp.
The product itself is not radical. What is interesting is the architectural claim underneath it. Microsoft is saying that Fabric should be the runtime for AI native applications, not just the analytics destination they write to. Stephanie Walter of HyperFRAME Research put it most plainly when she said the next phase of enterprise AI will be won by the platforms that can turn governed enterprise data into safe, operational applications. Rayfin is the mechanism by which Microsoft hopes to own that turn.
We see three reasons this matters for our roadmap. First, governance. Stewart Bond at IDC argues that the real value is inherited security, compliance and access policy from day one. Every backend deployed through Rayfin lands inside Fabric's permission model, lineage and audit, and the application data sits in the same governed estate that our analytics and AI workloads already consume. For CIOs and CDOs who have spent the last two years cataloging where agent built apps are storing customer records, that promise of a governed on ramp is hard to ignore. Ashish Chaturvedi at HFS Research went further, calling Rayfin an answer to the shadow IT explosion that coding agents have created, where every ungoverned app is a potential data silo and compliance liability waiting to surface.
Second, consolidation. David Linthicum points out that Rayfin collapses what is today separate app runtimes, data services, governance layers and custom integration code into one managed environment. For a large European retail group where store, supply chain and customer teams ship dozens of small operational apps a quarter, the integration tax between Fabric, App Service, Azure SQL and Power BI is real money. A single SDK that emits a backend already wired into the lakehouse, the semantic model and the access policy is a meaningful reduction in plumbing.
Third, the agentic angle. Rayfin is built to be authored by coding agents as much as by humans. That is the bet behind the SDK style: a deterministic, declarative API surface that an LLM can target reliably. If we want our internal agents to stand up customer facing micro backends without our platform team brokering every request, we need exactly this kind of guard railed runtime. The alternative, agents writing freeform code against arbitrary cloud services, is the failure mode CISOs are already pushing back against.
The skepticism is real and we should hold it. Walter notes that Microsoft has to prove the developer experience is lightweight, deployment is flexible, and governance benefits are strong enough to justify building inside Fabric rather than around it. Plenty of teams have evaluated Fabric, found the runtime opinionated and the cost model opaque, and stayed with their existing stack. A new SDK does not change that math by itself. We also need to watch how Rayfin interacts with non Microsoft data: pulling Snowflake or Databricks data into a Fabric backend still means moving or mirroring data, with the latency, cost and governance questions that brings.
For technology leaders at bruno.digital and our clients, the operator angle is straightforward. We should treat Rayfin as a serious option for any new AI native application that has a natural center of gravity inside the Microsoft estate, particularly where the data already lives in OneLake or Fabric warehouses. The pattern to pilot is a single tenant, low risk internal app, deployed through Rayfin, with explicit measurement of three things: developer hours saved versus the current Azure stack, time to first governed dataset in the lakehouse, and the number of net new IAM and DLP exceptions that the security team has to approve. Those metrics will tell us whether Rayfin really delivers the governance by default story or just shifts the work elsewhere.
We should also use this announcement to reopen the platform conversation. The market is now visibly converging on the idea that data platforms become application platforms, and not just for analytics agents. Snowflake has been laying down Cortex apps and the Horizon Context layer. Databricks has been extending Genie and the lakehouse runtime. SAP just acquired Dremio. Microsoft has Rayfin. Inside the next planning cycle we have to decide which one or two of these platforms we are willing to build agentic applications on, because trying to keep optionality across all four will produce the exact fragmentation Rayfin claims to solve. The right move is to pick the platform where the bulk of our regulated data already sits, run a sixty day agent app pilot on it, and publish the cost, latency and governance numbers internally before the next budget cycle.
