Azure/azure-sdk-for-java com.azure+azure-ai-agents_2.10.0

on GitHub

2.10.0 (2026-11-12)

Features Added

• Promoted voice-agent realtime WebSocket clients, session clients, persisted conversation clients, and core telephony
  clients to GA. Build the clients directly through AgentsClientBuilder; use allowPreview(true) when
  accessing capabilities that still require Foundry-Features: VoiceAgents=V1Preview.
• Added preview native GPT Live voice-agent definitions, WebSocket clients, and persisted conversation clients through
  AgentsClientBuilder.beta().buildBetaVoiceAgentsLiveWebSocketClient() and
  buildBetaVoiceAgentsLiveWebSocketAsyncClient(), with corresponding conversation-client factories.
  Live sessions support typed OpenAI ClientEvent sends and ServerEvent receives, raw JSON APIs, and Responses
  delegation through OpenAI ResponsesDelegationConfig. Live WebSocket clients automatically send the required
  preview header.
• Added GPT Live samples for multi-turn text, interactive text and audio, microphone/speaker streaming, Responses
  function tools, and persisted conversation history and audio. Persistence is opt-in through FOUNDRY_GPT_LIVE_STORE.
• Added preview voice-agent post-call delivery to Event Grid and routines, and inline subagents with handoff policies
  and runtime events.
• Added voice-agent system tools for DTMF, SMS, input collection, language detection, and transfer routing.
• Added preview SIP telephony bindings and telephony campaigns with recipient validation, scheduling, and long-running
  create, cancel, and delete operations on BetaVoiceAgentsTelephonyClient and its asynchronous counterpart.
• Added typed RealtimeResponseInputItem and RealtimeResponseInputContent factories for heterogeneous realtime
  response inputs and user-message content.
• Added skill optimization through SkillOptimizationConfiguration, skill references, validation and execution
  results, and resolved candidate context. Trace-based optimization supports Application Insights connections,
  fixed time windows, explicit trace IDs, and evaluator generation.
• Added optimization candidate files, errors, and execution summaries, and downloadOptimizationJobFile on
  AgentsClient and AgentsAsyncClient. Added skipped job and candidate statuses; optimization polling treats
  skipped jobs as unsuccessful terminal results.
• Added MCP toolbox request-header bindings.
• Added request-level model-routing mode and model-pool selection, adaptive session affinity, and preview
  required-parameter routing controls.

Breaking Changes

• Removed BetaAgentsClient and BetaAgentsAsyncClient. Use createAgentFromPrompt and
  createAgentFromPromptWithResponse on AgentsClient and AgentsAsyncClient.
• Replaced BetaVoiceAgentWebSocketClient, BetaVoiceAgentWebSocketAsyncClient, and their session clients with
  VoiceAgentsWebSocketClient, VoiceAgentsWebSocketAsyncClient, and the corresponding VoiceAgentsWebSocketSession*
  types. Build realtime clients directly with AgentsClientBuilder.buildVoiceAgentsWebSocketClient() or
  buildVoiceAgentsWebSocketAsyncClient() instead of beta(). The asynchronous session exposes closeAsync()
  instead of the previous blocking close().
• Replaced BetaVoiceAgentsConversationsClient and its asynchronous counterpart with VoiceAgentsConversationsClient
  and VoiceAgentsConversationsAsyncClient, built directly through AgentsClientBuilder.
• Moved binding, call, call-job, and transfer operations from BetaVoiceAgentsTelephonyClient and its asynchronous
  counterpart to VoiceAgentsTelephonyClient and VoiceAgentsTelephonyAsyncClient, built directly through
  AgentsClientBuilder. The beta telephony clients now expose campaign operations.
• Renamed VoiceAgentWebSocketConnectionOptions.getTransport() and setTransport(...) to
  getTransportType() and setTransportType(...).
• Replaced the numeric/string token-limit accessors on voice and realtime models with getMaxOutputTokens() and
  setMaxOutputTokens(VoiceAgentMaxOutputTokens), where writable. Use VoiceAgentMaxOutputTokens.of(count) for a
  numeric limit or VoiceAgentMaxOutputTokens.MAXIMUM for "inf".
• Replaced the per-variant tool-choice accessors on voice and realtime models with
  getToolChoice() and setToolChoice(VoiceAgentToolChoice), where writable. Use VoiceAgentToolChoice.ofMode,
  ofFunction, or ofMcp to create a choice. PromptAgentDefinition is unchanged.
• Changed VoiceAgentResponseCreateOptions input from List<RealtimeConversationItem> to
  List<RealtimeResponseInputItem>.
• Replaced local realtime transcription language and usage, session configuration, audio-format, turn-detection,
  and MCP tool/error models with OpenAI Java types. Notable changes include SessionUpdateEvent.Session,
  RealtimeAudioFormats, RealtimeMcpListTools.Tool, RealtimeMcpToolCall.Error, and the transcription event's
  getUsage() union. Foundry-specific session and turn-detection models remain available.
• Renamed session-created and session-updated event accessors from getSessionAsVoiceAgentSessionResponseConfiguration()
  to getSessionResponseConfiguration() and removed their public constructors.
• Renamed transcription getters from getLogprobs() to getLogProbs(), conversation-item object accessors to
  objectType, and voice audio/recording getBlobUri() accessors to getBlobUrl().
• Removed unit suffixes from duration-typed voice and telephony accessors. Audio item getStartOffsetMs() is now
  getOffset(). Transcription segment start/end values and server VAD prefix padding now use Duration.
• Renamed MemoryStoreDefaultOptions.getDefaultTtlSeconds() and setDefaultTtlSeconds(Duration) to
  getDefaultTtl() and setDefaultTtl(Duration), and RoutingTraceEntry.getLatencyMs() to getLatency().
  Wire values remain in seconds and milliseconds, respectively.
• Changed voice tool, greeting, interim-response, engine, error, and audio-timestamp type properties from strings to
  typed enums. Recording channel roles now use VoiceAudioRole.
• Changed VoiceAgentSubagentConfiguration to a polymorphic base. Use VoiceAgentDelegateSubagentConfiguration for
  delegate subagents or the preview VoiceAgentInlineSubagentConfiguration for inline subagents.
• Replaced telephony binding creation-input models with writable concrete binding models and
  CreateTelephonyCallJobInput with TelephonyCallJobCreate. Call sources and call/transfer destinations now use
  TelephonyAddress variants, and fixed-interval retries use TelephonyFixedIntervalRetryPolicy. Removed
  TelephonyCallJob.getObjectType() and binding-status accessors/filtering; added allowed calling numbers to binding
  updates.
• Replaced SessionAffinityRequestMode and SessionAffinityMode with SessionAffinityRequestStrategy and
  SessionAffinityStrategy. SessionAffinityDetails.getMode() is now getStrategy(); request mode accessors use
  SessionAffinityRequestStrategy. Removed SessionAffinitySource.CONVERSATION and PREVIOUS_RESPONSE_ID and added
  DERIVED_FINGERPRINT. The documented default affinity strategy is now adaptive rather than sticky.
• Hosted-agent A2A passthrough requires Foundry-Features: A2AAgents=V1Preview when creating or modifying a definition
  declaring the a2a protocol. allowPreview(true) adds the opt-in; HostedAgentDefinition.getProtocolVersions() and
  setProtocolVersions(...) are now marked preview.

Bugs Fixed

• Preserved exact numeric values when round-tripping agent structured input defaults and schemas, telephony
  structured inputs, and response structured inputs through BinaryData.
• Fixed OpenAI model JSON conversion to use an explicit Jackson serializer, preserving explicit nulls without
  depending on the application's Azure JSON serializer provider.
• Removed query values from WebSocket proxy-failure diagnostics to avoid logging structured inputs.

Other Changes

• Regenerated the client with @azure-tools/openai-typespec version 1.34.1.
• Updated openai-java from 4.69.0 to 4.99.0.