Azure/azure-sdk-for-java com.azure+azure-ai-agents_2.3.0

on GitHub

2.3.0 (2026-08-06)

Features Added

• Added AzureUserSecurityContext and AzureCreateResponseOptions.setUserSecurityContext(...) for supplying application and end-user context to Microsoft Defender for Cloud when creating an Azure response. See Protecting AI applications for details.
• Added AgentIdentityStatus (values ACTIVE and DISABLED) and AgentIdentity.getStatus() to expose the status of an agent's instance identity or blueprint identity.
• Added an allowedCallers property to many tool models, backed by the new CallableToolAllowedCaller enum (DIRECT, PROGRAMMATIC), so callers can restrict how a tool may be invoked. Supported types include ApplyPatchToolParameter, CodeInterpreterTool, CodeInterpreterToolboxTool, CustomToolParameter, FunctionShellToolParameter, FunctionTool, McpTool, and McpToolboxTool.
• Added programmatic tool calling support: new ProgrammaticToolCallingParameter tool and ToolType.PROGRAMMATIC_TOOL_CALLING value.
• Toolbox Search graduated to general availability. Added ToolSearchToolboxTool and ToolboxToolType.TOOLBOX_SEARCH for the GA tool; the preview ToolboxSearchPreviewToolboxTool (toolbox_search_preview) is retained alongside for backward compatibility.
• Added McpTool.setTunnelId(...) / getTunnelId() and the same on McpToolboxTool to route MCP calls through a Secure MCP Tunnel as an alternative to serverUrl or connectorId.
• Added FunctionTool.setOutputSchema(...) / getOutputSchema() for declaring the JSON schema of a function tool's return payload.
• Added OptimizationOptions.setMaxStalls(...) / getMaxStalls() for configuring early-stop behavior of preview agent-optimization jobs.
• Added ResponseUsageInputTokensDetails.getCacheWriteTokens() to expose the number of tokens written to the prompt cache in a response.

Breaking Changes

• BetaAgentsClient.createOptimizationJob(OptimizationJob) and createOptimizationJob(OptimizationJob, String) were replaced by long-running operations beginCreateOptimizationJob(OptimizationJob) and beginCreateOptimizationJob(OptimizationJob, String), which return SyncPoller<OptimizationJob, OptimizationJobResult> instead of OptimizationJob. The corresponding BetaAgentsAsyncClient methods now return PollerFlux<OptimizationJob, OptimizationJobResult>. The protocol method was likewise replaced by beginCreateOptimizationJob(BinaryData, RequestOptions) returning SyncPoller<BinaryData, BinaryData> / PollerFlux<BinaryData, BinaryData>.
• ResponseUsageInputTokensDetails now also carries a cacheWriteTokens field. Deserialized instances continue to work, but any code depending on the previous single-field structure should be updated to read the new value.

Other Changes

• Refined preview annotations. AgentDefinition and CreateAgentVersionInput are no longer marked @Beta at the class level; instead the preview-only fields (for example, CreateAgentVersionInput.getDefinition() and CreateAgentVersionInput.isDraft()) are individually annotated. Additional preview tool classes such as A2APreviewTool, BingCustomSearchPreviewTool, BrowserAutomationPreviewTool, FabricIqPreviewTool, MemorySearchPreviewTool, MicrosoftFabricPreviewTool, SharepointPreviewTool, WorkIqPreviewTool, and their toolbox counterparts are now marked @Beta so preview surface area is explicit in generated API docs.
• Reworked internal LRO polling-strategy customizations so that both BetaMemoryStoresClient and BetaAgentsClient long-running operations inherit the per-client Foundry-Features value from the builder pipeline instead of a hardcoded polling-only header. This lets each Beta sub-client's poll GETs carry the same preview opt-ins as its initial request.
• Regenerated client from the updated TypeSpec specification.
• Updated version of openai client library to 4.45.0.