Azure/azure-sdk-for-java com.azure+azure-ai-agents_2.7.0

on GitHub

2.7.0 (2026-10-06)

Features Added

• Browser automation graduated to general availability. Added BrowserAutomationTool and
  BrowserAutomationToolboxTool with the browser_automation discriminator, plus
  ToolboxToolType.BROWSER_AUTOMATION. The existing BrowserAutomationPreviewTool and
  BrowserAutomationPreviewToolboxTool remain available.
• Added estimateOptimizationJob, listOptimizationCandidates, getOptimizationCandidate, and
  promoteOptimizationCandidate to AgentsClient and AgentsAsyncClient, with corresponding protocol methods.
  Estimates expose call-count and cost bands by stage; candidate listing supports pagination and optional mutation
  expansion through AgentOptimizationCandidateExpand.
• Added typed agent-optimization configuration for baseline instructions, models, skills, and OpenAI function tools;
  model search spaces; evaluator initialization parameters; training and validation evaluation sets; and concurrent
  agent-run limits. AgentOptimizationConfiguration supports agent optimization, while
  PromptOptimizationConfiguration supports prompt optimization with steering-prompt context.
• Added target-completion and user-conversation-simulation evaluation sets with inline or registered-dataset sources.
  UserConversationSimulationConfiguration supports turn limits, conversation repetitions, audio effects, and user
  interruption behavior. EvaluationModelConfiguration supports text and voice model selection and
  ModelSamplingParameters, including a Long completion-token limit.
• Added FunctionTool.setAsync(...) and CustomToolParameter.setAsync(...) to configure asynchronous tool execution.
• Added ImageGenToolModel values for gpt-image-2, gpt-image-2.5-sunburst, and gpt-image-2.5-flare, including
  their dated model versions.
• Added ApiError.getMisalignment() to expose structured misalignment details using the OpenAI
  ErrorObject.Misalignment model.
• Added VoiceAgentWebSocketConnectionOptions.setTransport(...) and getTransport() to select
  VoiceAgentTransport.WEBSOCKET (the default) or VoiceAgentTransport.WEBRTC. WebRTC mode uses the WebSocket for
  signaling; the SDK does not implement a WebRTC peer connection or media transport.

Breaking Changes

• Moved agent-optimization job operations from BetaAgentsClient / BetaAgentsAsyncClient to
  AgentsClient / AgentsAsyncClient. Optimization no longer requires the AgentsOptimization=V2Preview header.
  BetaAgentsClient and BetaAgentsAsyncClient retain preview agent-generation operations.
• Replaced the no-argument AgentOptimizationJob constructor and getInputs() / setInputs(...) with
  AgentOptimizationJob(AgentOptimizationModelConfiguration, AgentOptimizationConfigurationBase) and top-level
  configuration properties. Use setTargetConfiguration(...) with
  AgentOptimizationFoundryAgentTargetConfiguration to identify a registered Foundry agent.
• Replaced the old optimization input hierarchy: AgentOptimizationJobInputs, AgentOptimizationOptions,
  OptimizedAgentIdentifier, AgentOptimizationDatasetInput, AgentOptimizationDatasetInputType,
  AgentOptimizationInlineDatasetInput, AgentOptimizationReferenceDatasetInput, AgentOptimizationDatasetItem,
  AgentOptimizationDatasetCriterion, and EvaluationLevel were removed. Use the new optimization configuration,
  evaluation-set, data-source, and test-case models instead.
• Optimization job listing now returns AgentOptimizationJob instead of AgentOptimizationJobListItem.
  AgentOptimizationJobProgress and AgentOptimizationJob.getProgress() were removed; use getRunDuration() and
  the job result's candidate summary. AgentOptimizationJobResult.getBaseline(), getBest(), and getCandidates()
  were replaced by getCandidateSummary(), with separate token-usage, latency, and termination-reason accessors.
  Retrieve candidates through listOptimizationCandidates(...).
• AgentOptimizationCandidate now exposes typed output and nested evaluation details instead of top-level mutations,
  scores, token averages, and evaluation IDs. Use getOutput() and getEvaluation(). PromotionInfo was replaced by
  AgentOptimizationCandidatePromotionInfo, which exposes the promoted agent through an AgentReference.
• RealtimeSessionUpdateEvent now has a single constructor accepting RealtimeSessionConfigurationBase and a public
  getSession() accessor. The constructors accepting RealtimeSessionConfiguration or
  RealtimeTranscriptionSessionConfiguration and their corresponding getSessionAs...() accessors were removed.
  Pass either concrete session subtype to the new constructor and inspect the subtype returned by getSession().
• Changed the constructors of MemoryItem and MemoryStoreDefinition from public to protected. Construct their
  concrete subtypes instead; deserialization of unknown discriminator values remains supported.
• Removed the unused public VoiceAgentFunctionToolType enum.

Bugs Fixed

• Fixed OpenAI clients built from AgentsClientBuilder to acquire bearer tokens through the Azure HTTP pipeline.
  Asynchronous clients now use asynchronous token acquisition instead of calling TokenCredential.getTokenSync(...);
  synchronous clients continue to use synchronous token acquisition.

Other Changes

• Added agent-endpoint configuration samples and optimization samples covering estimates, candidate inspection and
  promotion, job management, cancellation, and advanced synchronous and asynchronous polling.
• Updated conversation, tool, and hosted-agent samples, including agent-endpoint routing, asynchronous response
  handling, session-log streaming, file-search uploads, and OpenAPI project-connection authentication.
• Regenerated the client from the updated TypeSpec specification.

Dependency Updates

• Updated openai-java from 4.45.0 to 4.69.0.
• Upgraded Netty dependencies from 4.1.137.Final to 4.1.138.Final.