How Microsoft wants Fabric to govern enterprise AI agents

Microsoft wants Fabric to govern agentic work

Microsoft wants enterprise technology leaders to give Fabric a larger role in agentic AI, where software agents carry out tasks under organizational controls. Its announcements at the European Microsoft Fabric + SQL Community Conference in Barcelona extend the data platform into delegated engineering, application development, operational supervision, database administration, and sharing between organizations. For CIOs and engineering leaders, the decision is how much execution and oversight to place within that platform.

Microsoft’s case starts with preparing organizational data for agents to understand and use. “We’re going to talk a lot about the notion of how you prepare your data estate to really provide the context for AI agents to help you transform the organization,” Amir Netz, CTO of Azure Data at Microsoft, told CIO.com in the week before the announcements. Microsoft sees enterprise agentic transformation as already underway and benefits commercially when organizations choose Fabric for that work. Netz’s account explains the role Microsoft wants its platform to play.

That role connects the meaning of data with control over its use. “Fabric is becoming the center point for everything data at Microsoft and everything that’s context based on data,” Netz said. The announcements turn that ambition into several responsibilities: engineers assign bounded tasks to agents, developers build applications alongside governed data, and administrators oversee resources that can remain outside Fabric. Each requires a decision about what the platform should execute or manage.

Engineering agents need clear assignments and boundaries

Microsoft’s engineering changes start with technology from Osmos, the AI-based data engineering company it acquired in January. That technology underpins Fabric’s new agentic engineering features, which use agents to prepare raw data for analytics and AI in OneLake, Fabric’s central data lake. Engineers specify the required result and the limits within which the agent can work.

With those instructions set, the agent performs the task autonomously, checks its results, and revises its approach. Validation is part of execution: the agent assesses whether its work meets the assignment before refining it. So an engineering team must define both the desired data asset and the constraints on producing it. Evaluation can then test whether the agent completes the assignment accurately within those limits.

Netz attributes the system’s accuracy to the Osmos technology. “They have this secret sauce of how to get really high accuracy from agents in very complex data engineering projects that you can’t just achieve by telling Claude to do it for you,” he said. Since Claude is a competing option for this work, this is a competitive claim from Microsoft. His argument puts the execution, checking, and revision process at the center of complex engineering work.

Netz extends the comparison to other commercial systems: “No matter how well your prompt is constructed, it’s not going to be good enough. The accuracy level we’re getting from this Osmos technology is significantly higher than any other commercially available agentic system from the best frontier labs.” Teams considering delegation need to test whether that accuracy meets their engineering requirements. Representative assignments should exercise the full process, including the agent’s validation and revision of its output.

Application development brings hosting and access decisions

Delegated engineering prepares data for use; application development determines how users and software act on it. Microsoft’s Fabric Apps supports that next step through Rayfin, the open-source software development kit and command-line interface unveiled at Microsoft Build 2026 in June. A software development kit supplies tools for building applications, while a command-line interface lets developers operate tools through typed commands. Microsoft aims to bring AI-assisted application development alongside data already governed in Fabric.

Within that environment, Fabric Apps supports data connectivity, transactional workloads, backend logic, storage, and policy-controlled access. Transactions perform application operations, backend logic determines application behavior, and storage retains data. Access policies govern who can use the relevant data and functions. Together, these capabilities give developers the components to connect a user’s action to work on governed data.

Netz says this scope covers the application’s interface and the processes behind it. “It’s not just the data itself,” Netz said. “You can build the entire app, including the web pages, the business processes, the functions, and you can build it end-to-end in Fabric. You don’t need to deploy containers. It’s fully hosted, fully managed, fully governed, only Fabric.” Microsoft’s commercial proposition therefore includes running the application after developers build it.

This changes the handoff between development and operations because teams can build pages, processes, and functions without separately deploying containers, the software packages commonly used to run applications. Fabric supplies the hosting, management, and governance for this environment. Engineering leads must decide whether that arrangement fits their application’s execution, storage, and access requirements. Agent accuracy answers a separate question about the quality of delegated work; application management sets the conditions under which that work runs.

Production visibility must follow agents into code

Once an application is running, its operators need to understand how its actions produce a result or a failure. Microsoft’s new Observability in Fabric capability addresses this need: observability means being able to inspect a system’s activity well enough to understand its behavior and investigate problems. Microsoft positions it as support for proactive operations, allowing teams to investigate developing issues and initiate responses.

Investigation starts with Monitor Hub, which combines views of jobs, capacities, workspaces, and workloads. These cover individual execution tasks, their available computing resources, the areas where teams organize work, and the work those resources support. Microsoft brings the views together so operators can examine a job within its surrounding environment. That context helps connect an application’s behavior to the resources and activities involved.

From that view, operations agents and Activator help teams investigate issues, use AI to identify root causes, and automate responses. The sequence matters because an automated response changes the running system based on the investigation’s findings. Teams should therefore examine how an observed event leads to a diagnosis and then an action. Operators need enough visibility to supervise the whole process.

Netz says this visibility supports both development and production: “Observability is required for teams to both debug their projects and manage projects after they’re released in production.” Before release, engineers investigate why a project behaves differently from their intent. After release, operators investigate behavior while managing a system already in use. Following execution supports both responsibilities, although the consequences of intervention differ.

For agentic applications, following execution also means tracing the code an agent invokes. “It’s spanning both agents and code because you have agents that are going to call code, and you need to have the entire journey of the logic, and the execution from end-to-end,” Netz said. An investigation must connect an agent’s activity with the code it calls and the resulting outcome. Engineering teams can test that visibility by tracing representative execution paths, including the handoff from an agent to code.

Database oversight extends across environments

Application execution is one operational responsibility; the databases supporting an organization create another. Microsoft’s Database Hub in Fabric, available in public preview, offers a central control plane, meaning a shared management interface for overseeing resources. It covers SQL Server, Azure SQL, PostgreSQL, Cosmos DB, and SQL databases in Fabric. Microsoft proposes using the interface to centralize monitoring and governance across these offerings.

This administrative reach extends to databases in Azure and on-premises environments, where organizations run systems on their own infrastructure. Administrators can discover, monitor, manage, and administer those databases with agent assistance. So the location of the management interface is a separate decision from the location of the databases it oversees. Fabric Apps concerns application construction and hosting, while Database Hub gives Fabric administrative reach into other environments.

Netz presents this reach as an extension of Fabric’s existing purpose: “Fabric has always been about unifying all the data services, unifying all the data, and unifying users in one place. Now it unifies the entire inventory of databases, both in Azure and on-premises, in one place. You can find all your databases, monitor them, manage them, administrate them, and do it all with the help of an agent.” His description places database administration within Microsoft’s broader effort to expand Fabric’s role.

For an organization considering that role, the immediate test is whether Database Hub covers the databases and administrative tasks its teams need to manage. Administrators should compare the five named database offerings with their inventory and examine how agent assistance fits their responsibilities. Public-preview status also matters when deciding which administrative work to place under the new interface.

Cross-organization agents need shared business context

Oversight across technical environments leads to a further boundary: work involving separate organizations. Microsoft’s IQ Share addresses this by allowing organizations to share data securely with its business context while retaining governance. Recipients can include agents outside the organization holding the data. Those agents need both permission to use the information and enough context to interpret it for the intended work.

IQ Share proposes carrying that context with the shared data. Netz describes a two-party scenario: “The agents work on the same copy across the two organizations and the context comes with it so agents on both sides can really understand the meaning of the data.” The common copy supplies a shared reference, while the accompanying context explains what the information means. Microsoft’s proposition is that agents on each side can use both under the governance attached to the exchange.

For technology and data leaders, a sharing decision therefore depends on how the receiving agent interprets the information. An evaluation should follow the intended joint task and test whether the shared context preserves the organizational meaning needed to perform it. Agents in both organizations must be able to carry out that work using the common data while observing the conditions governing its use.

Key takeaways for decision-makers

• Test delegated engineering on real assignments: Engineering teams evaluating Fabric’s Osmos-based agents need to define task boundaries and test the full execution, validation, and revision process. Microsoft’s accuracy claims require verification against representative workloads.
• Evaluate application hosting alongside development: Fabric Apps brings application logic, storage, and policy-controlled access into a Microsoft-managed environment. Engineering leads need to assess whether that arrangement fits their execution, access, and operational requirements.
• Trace agents through code and automated responses: Operations teams evaluating Fabric observability need to follow agent activity into code execution and connect diagnoses to resulting actions. Testing those paths establishes whether operators have enough visibility to supervise production work.
• Match database oversight to administrative needs: Database administrators evaluating Database Hub’s public preview need to compare its supported offerings and management functions with their inventory. Fabric can centralize oversight while databases remain in Azure or on premises.
• Validate context across organizational boundaries: Data owners evaluating IQ Share need to test whether external agents interpret shared data correctly and observe its governance conditions. A representative joint task can establish whether the accompanying business context supports the intended work.