Azure/azure-sdk-for-python azure-ai-projects_2.11.0

on GitHub

2.11.0 (2026-11-13)

Features Added

• Voice Agents are generally available through .voice_agents.realtime, .voice_agents.conversations, and the existing binding, call, transfer-target, and call-job operations on .voice_agents.telephony. Agent generation is now available through .agents.create_from_prompt(). GPT Live, inline subagent handoff, SIP, post-call delivery, and campaigns remain in preview.
• External Agents are generally available. Creating agents with ExternalAgentDefinition no longer requires allow_preview=True.
• Agent Insights are generally available through .agent_insight_monitors, with AgentInsightPriorities, AgentInsightPrioritiesRead, AgentInsightRunInputs, and AgentInsightSummary.
• Added skill optimization through SkillOptimizationConfiguration, with trace-based learning, skill references, optional evaluation, and typed execution, validation, and result models. Added AgentTracesOptimizationContext, SkillReferencesOptimizationContext, and SkillOptimizationEvaluationConfiguration.
• Added optimization trace selection through AgentOptimizationFixedTraceWindow, AgentOptimizationTraceIdsTraceWindow, and AgentOptimizationAppInsightsTraceConnectionConfiguration.
• Added .agents.download_optimization_job_file() and AgentOptimizationCandidateFile for downloading optimization artifacts. Added optional error, execution_summary, files, and resolved_context properties on AgentOptimizationCandidate, and AgentOptimizationCandidateStatus.SKIPPED.
• Added shared JobStatus.SKIPPED for Agent Insights runs, agent optimization, data generation, evaluator generation, and telephony campaign operations. Agent optimization job and Agent Insights run status filters also accept this value.
• Added quality and safety rubric generation through EvaluatorGenerationCategory, Boolean and Likert scoring through RubricScoring and RubricScoringMode, and optional Dimension.allowed and read-only Dimension.generated_weight. Dimension.weight is now optional.
• Added optional category, dimensions, pass_threshold, and scoring properties on EvaluatorGenerationInputs and EvaluatorGenerationJob. EvaluatorGenerationInputs.model is now optional for safety generation. Added optional TracesEvaluatorGenerationJobSource.trace_ids and RubricBasedEvaluatorDefinition.scoring.
• Added preview .beta.evaluators.list_diagnostics() for evaluator publication lineage and diagnostics, with actor, change, and health metric models.
• Added optional EndpointBasedEvaluatorDefinition.schema_version and EndpointEvaluatorSchemaVersion for selecting an endpoint evaluator contract version.
• Added scheduled dataset and trace evaluations through DatasetEvaluationRuleAction and TraceEvaluationRuleAction. EvaluationRule now exposes optional schedule, provisioning_status, and error properties, and EvaluationRuleEventType includes TIMER_TRIGGER.
• Added typed evaluation input sources for JSONL, CSV, stored completions, target completions, conversation simulation, and agent traces, including CreateEvalJsonlRunDataSource, CsvEvalRunDataSource, TargetCompletionEvalRunDataSource, UserConversationSimulationEvalRunDataSource, TraceSourceEvalRunDataSource, and TraceSourcePreviewEvalRunDataSource. Evaluation-rule annotations and the DatasetEvaluationSource and TraceEvaluationSource unions use these model names; wire discriminator values are unchanged.
• Added DataGenerationJobType.CONVERSATIONS, UserConversationSimulationDefaultConfiguration, and optional TracesDataGenerationJobConfiguration.generate_test_case_rubric. DataGenerationJobResult.warnings now exposes DataGenerationWarning and DataGenerationWarningCode.
• DailyRecurrenceSchedule.hours is now optional; when omitted, the service uses the current UTC hour at schedule creation.
• Added optional MCPToolboxTool.request_header_bindings, with MCPToolboxRequestHeaderBinding and MCPToolboxRequestHeaderBindingSource, for forwarding client-supplied headers to an MCP server.
• Added preview GPT Live sessions through .beta.voice_agents.live.connect() and VoiceAgentLiveDefinition, with typed configuration, session updates, text and audio input, lifecycle and usage events, and Responses backend tools.
• Added GPT Live delegation through VoiceAgentLiveResponsesDelegation, VoiceAgentLiveClientDelegation, and VoiceAgentLiveManagedAgentDelegation for Responses backends, application-managed work, and same-project prompt agents.
• Added .beta.voice_agents.live_conversations for managing stored Live conversations and retrieving conversation items and audio.
• Added optional preview VoiceAgentDefinition.post_call with VoiceAgentPostCallConfig, VoiceAgentEventGridPostCallDestination, and VoiceAgentRoutinePostCallDestination for post-call transcript and optional audio delivery.
• Added preview SIP telephony through TelephonyProvider.SIP, SipTelephonyBinding, and SipTelephonyBindingListItem. Added optional UpdateTelephonyBindingRequest.allowed_calling_numbers.
• Added .voice_agents.telephony.list_call_jobs() and .voice_agents.telephony.delete_call_job().
• Added preview outbound campaign operations on .beta.voice_agents.telephony, including long-running creation, cancellation, and deletion, listing and retrieval, validation errors, operation status, and ETag match conditions.
• Added campaign configuration, scheduling, and progress models, including TelephonyCampaign, TelephonyRecipient, ImmediateSchedule, AtTimeSchedule, PerRecipientSchedule, and TelephonyCampaignProgress. Added optional TelephonyCampaignDatasetReference.version and read-only TelephonyCallJob.campaign_name.
• Added preview inline Voice Agent subagents and handoff through VoiceAgentInlineSubagentConfig, InlineVoiceSubagent, and VoiceAgentHandoffPolicy, with handoff lifecycle events and optional VoiceAgentSessionResponseConfig.handoff.
• Added Voice Agent system tools VoiceAgentTransferSystemTool, VoiceAgentSendDtmfSystemTool, VoiceAgentSendSmsSystemTool, VoiceAgentCollectDtmfSystemTool, VoiceAgentCollectSmsSystemTool, VoiceAgentDigitRecognitionSystemTool, and VoiceAgentLanguageDetectionSystemTool, with transfer targets and routing models. Added optional VoiceAgentSystemTool.timeout.
• Added realtime system-tool items, argument-streaming events, and input-collection events through VoiceAgentSystemToolItem, VoiceAgentServerEventSystemToolArgumentsDelta, VoiceAgentServerEventSystemToolArgumentsDone, and VoiceAgentServerEventInputCollectionStatus.
• Added VoiceAgentBaseDefinition, VoiceAgentVoiceConfig, and VoiceAgentSubProtocol for shared Realtime and Live configuration, with voice-specific tool, greeting, conversation-engine, and event-discriminator enums.
• Added optional MisalignmentErrorDetailsResource.review_target and RealtimeResponseStatusDetailsError.message.
• Expanded known-value annotations on ImageGenTool.model for the gpt-image-2.5-sunburst and gpt-image-2.5-flare families and on ImageGenTool.quality for xhigh and max.
• Added explicit recursive typing for nested CompoundFilter instances in CompoundFilter.filters, while retaining support for arbitrary values.

Breaking Changes

Breaking changes in beta methods:

• Moved .beta.agents.create_from_prompt() to .agents.create_from_prompt(), and .beta.voice_agents.realtime, .beta.voice_agents.conversations, and the existing binding, call, transfer-target, and call-job operations on .beta.voice_agents.telephony to .voice_agents. GPT Live and campaign management remain under .beta.voice_agents.
• Renamed argument telephony_binding to binding on the promoted .voice_agents.telephony.create_binding() method. Its typed body now accepts TelephonyBinding instead of CreateTelephonyBindingRequest.
• Renamed argument body to job on the promoted .voice_agents.telephony.create_call_job() method. Its typed body now accepts TelephonyCallJobCreate instead of CreateTelephonyCallJobRequest.
• Removed the status filter from the promoted .voice_agents.telephony.list_bindings() method.
• Moved .beta.agent_insight_monitors to .agent_insight_monitors, and renamed BetaAgentInsightMonitorsOperations to AgentInsightMonitorsOperations.
• Agent Insights create(), update(), begin_create_run(), and update_insight() now accept flattened keyword arguments or JSON bodies instead of typed request models. Positional body arguments monitor, run, and update are now named body.
• Renamed Agent Insights list_insights() filters category, severity, and status to categories, severities, and statuses, respectively; they now accept lists instead of single values.

Breaking changes in beta classes:

• Renamed BetaRealtime, BetaRealtimeConnection, and BetaRealtimeConnectionManager to Realtime, RealtimeConnection, and RealtimeConnectionManager, respectively, with corresponding asynchronous class renames. The old exports were removed.
• Removed AgentInsightMonitorCreate, AgentInsightMonitorUpdate, AgentInsightRunCreate, and AgentInsightUpdate; use flattened operation arguments or JSON bodies. AgentInsightRun.inputs now uses AgentInsightRunInputs.
• Removed AgentInsightOverviewSource, AgentInsightsOverview, AgentInsightsOverviewOverride, and AgentInsightMonitor.overview. Use AgentInsightMonitor.agent_summary and priorities.
• Removed CreateTelephonyBindingRequest, CreateTeamsPhoneExtensibilityTelephonyBindingRequest, and CreateTwilioTelephonyBindingRequest. Use TelephonyBinding, TeamsPhoneExtensibilityTelephonyBinding, and TwilioTelephonyBinding, respectively.
• Renamed CreateTelephonyCallJobRequest to TelephonyCallJobCreate and TelephonyOutboundFixedIntervalRetryPolicy to FixedIntervalRetryPolicy.
• Removed TelephonyOutboundDestination and TelephonyOutboundDestinationType; use TelephonyAddress and TelephonyAddressType. Call-job source and destination now use TelephonyAddress; source no longer accepts a plain string.
• Removed TelephonyTransferDestination, PSTNTelephonyTransferDestination, SipTelephonyTransferDestination, TeamsTelephonyTransferDestination, and TelephonyTransferDestinationKind. TelephonyTransferTarget.destination now uses TelephonyAddress, with PhoneNumber, SipAddress, and TeamsAddress variants. The discriminator is type instead of kind, and the PSTN value is phone_number instead of pstn.
• Removed TelephonyBindingStatus and the status property from bindings, binding list items, their provider-specific subclasses, and UpdateTelephonyBindingRequest.
• TelephonyBinding.id and incoming_call_url are now read-only. Removed TelephonyCallJob.object; its service-managed identity, status, cancellation, retry, revision, and timestamp properties are now read-only.
• VoiceAgentSubagentConfig is now a polymorphic base requiring subagent_type instead of subagents. Use VoiceAgentDelegateSubagentConfig for referenced agents or VoiceAgentInlineSubagentConfig for inline subagents.
• Renamed VoiceAgentSubagent.invoke_timeout_seconds to invoke_timeout. The wire name is unchanged.
• VoiceAgentResponseCreateParams.input now uses dedicated response-input models instead of RealtimeConversationItem.
• Constructor argument timestamp_type on VoiceAgentServerEventResponseAudioTimestampDelta is now required; use VoiceAgentAudioTimestampType.WORD. Constructor arguments left and right on VoiceRecordingChannelLayout are now required; use VoiceAudioRole.USER and VoiceAudioRole.AGENT.
• Renamed VoiceAgentInputTranscriptionModel.GPT4_O_TRANSCRIBE, GPT4_O_MINI_TRANSCRIBE, GPT4_O_TRANSCRIBE_DIARIZE, and WHISPER1 to GPT_4_O_TRANSCRIBE, GPT_4_O_MINI_TRANSCRIBE, GPT_4_O_TRANSCRIBE_DIARIZE, and WHISPER_1, respectively. Wire values are unchanged.

Bugs Fixed

• Fixed the GPT Live sample helper to enforce the overall event-wait deadline when sending silence audio. Stalled audio sends are now cancelled and raise TimeoutError instead of hanging indefinitely.

Dependency update

• Added Python 3.15 to the package's supported-version classifiers. The minimum Python version remains 3.10.

Sample updates

• Added sample_voice_agent_campaign.py demonstrating future-scheduled outbound campaign creation, polling, cancellation, and deletion.
• Added sample_voice_agent_live_read_conversation.py demonstrating stored GPT Live conversation, transcript, and audio retrieval.
• Added sample_voice_agent_read_generated_audio.py demonstrating retrieval of generated audio from an interrupted Realtime reply.
• Added GPT Live examples demonstrating session lifecycle and settings updates, text and audio conversations, function tools, and same-project managed-agent delegation, with shared audio and event helpers.
• Added a Realtime interruption example demonstrating response cancellation and conversation-item truncation.
• Added sample_dataset_generation_job_traces_for_evaluation_with_test_case_rubric.py demonstrating evaluation-dataset generation with test-case rubric dimensions from agent conversation traces.
• Added sample_audio_conversation_evaluation.py demonstrating grading existing voice conversations with built-in audio evaluators.
• Added sample_voice_agent_simulation_evaluation.py demonstrating simulated voice conversations graded with audio evaluators.
• Added sample_dataset_evaluation_rule.py demonstrating scheduled evaluation against a fixed query dataset, and sample_trace_evaluation_rule.py demonstrating scheduled evaluation of Application Insights traces.
• Added sample_endpoint_evaluator.py demonstrating endpoint-evaluator registration through a project connection and inline-data evaluation, and sample_endpoint_evaluator_schema_0_0_1.py demonstrating the older endpoint contract. Removed the separate API-key and Entra ID samples after consolidating their scenarios.
• Updated rubric examples for Likert, Boolean, and safety scoring, pass thresholds, allowed dimension values, and generated weights.
• Updated Agent Insights examples for the GA client and flattened operation arguments, and External Agent and Voice Agent examples for GA paths and preview opt-in requirements.
• Updated optimization polling examples to handle skipped jobs as terminal outcomes with no optimization result.
• Updated Realtime audio, conversation, logging, and configuration examples. Consolidated the synchronous Realtime text-conversation example into the asynchronous example, moved the function-tool example to an asynchronous sample, and removed the separate basic asynchronous Voice Agent sample.
• Updated installation instructions in samples added or modified for this release to require azure-ai-projects>=2.11.0, and refreshed the datasets and evaluations sample READMEs.