Note
Access to this page requires authorization. You can try signing in or changing directories.
Access to this page requires authorization. You can try changing directories.
Every Fabric data agent runs on a runtime. The runtime determines the agent's core components: the orchestration, planning, and routing logic, along with the built-in query-generation tools that translate natural-language questions into queries against your data sources.
Fabric offers two runtimes:
- Standard runtime — the generally available (GA) runtime, optimized for stable, predictable behavior.
- Preview runtime — a runtime with the latest improvements and modifications to core components (for example, the built-in query-generation tools or the agent's routing logic), before those changes graduate to GA.
The runtime you choose determines how and when changes to the agent's core components reach your agent. It doesn't determine which data sources or preview features you can add.
Important
Model upgrades to the underlying large language model (LLM) are applied consistently across both the standard and preview runtimes. Runtime selection doesn't control which model the data agent uses.
Standard runtime
The standard runtime is the GA runtime and is the default for new data agents. It contains the generally available implementations of the agent's core components, including:
- Built-in query-generation tools that translate natural-language questions into queries against supported data sources, including Lakehouse, Data Warehouse, SQL Database, Mirrored Database, Eventhouse KQL Database, and Power BI semantic model.
- The core agent orchestration, planning, and routing logic.
Updates to the standard runtime are infrequent. Changes land here only after they pass validation in the preview runtime and graduate to GA. Choose the standard runtime for production data agents where consistent behavior and minimal churn between releases are important.
Preview runtime
The preview runtime contains the latest improvements to the built-in tools and orchestration before they graduate to GA. These updates can include:
- New built-in tools that aren't yet available on the standard runtime.
- Modifications and quality improvements to existing built-in tools (for example, updates to how the agent generates or validates queries for a given data source).
- Changes to the agent's core orchestration, planning, or routing logic.
Updates to the preview runtime are more frequent, and behavior can change between releases. Choose the preview runtime when you want to evaluate upcoming changes, validate that your agents continue to work as expected, or provide feedback on improvements before they reach GA.
What's currently in the preview runtime
The following updates are included in the preview runtime today. This list changes as new improvements land and as existing ones graduate to GA.
| Improvement | Feature | Description |
|---|---|---|
| Better example query following | Advanced NL2SQL | NL2SQL adheres more closely to the patterns shown in your example query library, instead of adding extra logic or constraints that weren't in the examples. |
| Filter value substitution | Advanced NL2SQL | NL2SQL reasons through implied filter values and substitutes the correct ones, including when multiple categorical or boolean filters are implied rather than explicitly stated. |
| Ambiguity handling | Advanced NL2SQL | NL2SQL detects ambiguous questions and asks a clarifying question before generating SQL, instead of committing to an assumption that may produce the wrong answer. |
You can also switch a data agent to the preview runtime programmatically by using the Fabric data agent Python SDK:
data_agent = FabricDataAgentManagement(data_agent_name)
# Switch to the preview runtime
data_agent.update_configuration(enable_preview_features=True)
config = data_agent.get_configuration()
print(f"Preview runtime enabled: {config.enable_experimental_features}")
# Switch back to the standard runtime
data_agent.update_configuration(enable_preview_features=False)
config = data_agent.get_configuration()
print(f"Preview runtime enabled: {config.enable_experimental_features}")
Note
The runtime that a published data agent uses is set by the agent's configuration at the time of publishing. If you publish a data agent while it's configured to use the preview runtime, the published version continues to run on the preview runtime until you republish the agent with a different runtime selection.