Skip to content
Agents tracked: 501 Releases (7d): 777 Agent status: no reported issuesSince last limit reset: Claude 10d · Codex 8d For AI agents: Markdown Updated Nov 21, 2026

Agent Development Kit (ADK) changelog: what's new each month

Every stable Agent Development Kit (ADK) release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 12 months covered; the current month updates daily.

November 2026 · so far

3 releases: 2.14.0 → 2.15.0

November brought stable execution environments, improved file handling for large data, and new integration capabilities. The month focused on performance, evaluation rigor, and expanding agent tooling.

Highlights

  • Execution environments (LocalEnvironment, E2BEnvironment, DaytonaEnvironment) are now stable and stream files in chunks to avoid loading entire files into memory.
  • New GoogleMapsToolset lets agents verify addresses, reverse-geocode coordinates, and fetch Street View imagery directly.
  • BigQueryToolset now accepts search scope parameters to answer analytics questions without knowing exact table references upfront.
  • Multi-run evaluation now distinguishes between pass-at-k (solving once) and pass-hat-k (solving reliably every time).
  • FsspecArtifactService unifies artifact storage across local development and cloud stores (Google Cloud Storage, S3, file, memory).

New

  • GoogleMapsToolset with get_street_view_image, validate_address, and reverse_geocode functions.
  • BigQueryToolset search_scope parameter for analytics queries without predefined table references.
  • BufferedSessionService (experimental) buffers session events in memory to keep writes off the critical path.
  • FsspecArtifactService for unified artifact storage across local and cloud object stores.
  • BaseEnvironment.read_file_stream for streaming files in chunks across all environment types.
  • ContainerCodeExecutor support for input and output files.
  • AgentEngineEnvironment timeout detection for commands exceeding 570 seconds.
  • Runtime-built agents support in AgentEvaluator.

Improved

  • MultiRunConfig and --parallelism flag enable concurrent evaluation runs with combined scoring.
  • ReadFileTool can now read line ranges without loading entire files into memory.
  • RunSkillScriptTool now validates paths and arguments instead of accepting raw shell command strings.
  • BigQueryToolset ask_data_insights accepts includeProjectIds and includeDatasetReferences parameters.
  • --eval_storage_uri stores evaluation files outside read-only agents_dir.
  • Agent Engine now reports multi-run evaluation results showing both pass@k and pass^k metrics.

Fixed

  • FastAPI native tracing disabled on Agent Engine to reduce overhead.

October 2026

3 releases: 2.11.0 → 2.13.0

October brought significant performance improvements, stability enhancements, and expanded deployment capabilities to ADK. The month featured 12 experimental features graduating to stable status alongside new tools for voice interactions and workflow control.

Action needed
  • Telemetry defaults to OpenTelemetry GenAI v2 (otel_semconv_1_44); set ADK_TELEMETRY_SCHEMA_VERSION_OPT_IN=otel_semconv_1_36 to use legacy v1 schema.
  • call_llm span deprecated in favour of OpenTelemetry 1.44 GenAI semconv.

Highlights

  • Live audio now supports real-time voice interactions with barge-in, Gemini speech-to-text, and text-to-speech integration.
  • Performance optimized: 40% faster time to first model request in long sessions and up to 8x faster execution turns for string-model agents.
  • Execution can now be cancelled gracefully via abort signals, with automatic cancellation when SSE clients disconnect.
  • Workflows gain human confirmation requirements and tool output filtering before model processing.
  • MongoDB support for persistent sessions shared across runner replicas, plus server-side embedding without API calls.

New

  • Execution cancellation via abort_signal for Runner, Workflow, and nodes with automatic handling of SSE client disconnection.
  • Tool confirmation in workflows pauses for user approval through RequestInput instead of passing errors downstream.
  • ModelConsultTool allows agents to consult other models mid-task with per-turn and per-session budget limits.
  • SQLite memory service stores agent memory in a local SQLite database using sqlite:// URI.
  • MCP SDK 2.x support connects to MCP servers over the modern protocol through an opt-in path.
  • TypesafeClassifier and ClassifierNode route workflows using TypeSafe's Jev models with calibrated probabilities in single forward pass.
  • Live audio interactions with barge-in support and Gemini speech-to-text and text-to-speech integration.
  • MongoDbSessionService persists sessions in MongoDB and shares them across runner replicas.

Improved

  • Tool output interception and filtering via after_tool_callback before returning to the model.
  • Workflow node execution now requires human confirmation through require_confirmation.
  • Python 3.15 installation and runtime support.
  • Agent deployment to Agent Engine using dedicated custom service accounts.
  • MongoDbToolset deployment to Agent Engine without packaging connection strings, with server-side embedding via Atlas Automated Embedding.
  • Performance reduced event-loading overhead alongside faster execution turns.
  • Server commands now support avatar configuration and max LLM calls settings.
  • Added state_file option to adk run command for improved session management.

Fixed

  • Telemetry schema updated to OpenTelemetry GenAI v2 (otel_semconv_1_44) with legacy v1 schema available via environment variable.
  • Security enhancements with skill-level security support added to AgentCardBuilder.

September 2026

4 releases: 2.9.0 → 2.10.0

September brought significant improvements to agent reliability, voice capabilities, and workflow management, along with new tools for vector search and better model integration. The release cycle included refinements to safety, telemetry, and cost optimization features.

Action needed
  • GCS tools refuse local file access entirely when localfileroot setting is absent (previously accepted and discarded).
  • InMemorySessionService now raises SessionNotFoundError for events on non-existent sessions (previously accepted silently).
  • Workflow nodes that fail now re-run on workflow resume; ensure node bodies are idempotent to avoid unexpected behavior.

Highlights

  • Automatic model failover now protects agent uptime by switching to backup models when the primary encounters errors.
  • New LiveKit runner enables building rich voice and telephony agents.
  • YAML-based workflow configuration makes graph authoring simpler and more declarative.
  • MongoDB toolset adds high-performance vector and hybrid search capabilities.
  • New evaluation metrics track agent costs and latency through duration, token consumption, and model call counts.

New

  • LiveKit runner for voice and telephony agent development.
  • YAML configuration support for authoring and loading ADK 2.0 workflow graphs.
  • MCP SDK 2.x support with full backward compatibility for 1.x servers.
  • MongoDB database integration toolset for vector and hybrid searches.
  • Ephemeral skill lifecycles and active skill limits for dynamic resource control (experimental).
  • Evaluation metrics for tracking agent duration, token consumption, and model call counts.
  • Visible thoughts access for Claude models during adaptive thinking tasks.
  • Automated parameter adaptation for OpenAI reasoning models with precise reasoning token reporting.

Improved

  • Workflow node resumption now re-runs failed nodes instead of replaying them as completed.
  • GCS tools restricted to read and write only within the localfileroot directory.
  • OpenTelemetry event names now preserved on platforms outside Agent Engine for consistent tracking.
  • BigQuery protected write mode validates dry runs before executing multi-statement scripts and other statements.
  • BigQuery protected mode now keeps sessions in process memory instead of session state to prevent temporary table persistence across restarts.
  • InMemorySessionService now raises SessionNotFoundError for events appended to non-existent sessions instead of silently discarding them.

Fixed

  • OpenAIResponsesLlm now warns when thinkingconfig is used and directs users to use OpenAIGenerateContentConfig.effort instead.
  • AgentEvaluator.evaluate and evaluateevalset now raise ValueError when no eval cases are evaluated instead of passing silently.

August 2026

8 releases: 2.6.2 → 1.39.1

August brought major improvements to model capability handling, tool response formatting, and conversation preservation across ADK versions. The month also saw performance optimizations and expanded integration support.

Action needed
  • pyarrow moved from gcp extra to new bigquery-analytics extra; users of BigQueryAgentAnalyticsPlugin must install the bigquery-analytics extra explicitly.
  • gcloud beta flag now required for sandbox-launcher gate functionality.

Highlights

  • Models now declare their own capabilities instead of ADK inferring support from model IDs.
  • Tools can return media directly to models including images on Gemini, Anthropic, LiteLLM, Apigee and OCI.
  • Conversation history now preserves thought signatures, server-side tool call parts, and parallel function call results.
  • Performance improvements reduced import overhead and eliminated inefficiencies in request handling and streaming.
  • Jinja2 templates now supported for instruction injection with access to session state and artifacts.

New

  • Jinja2 instruction templates with conditional and loop support over session state.
  • Pre-configured client support for the labs OpenAI model.
  • Native task mode support for RemoteA2aAgent.
  • ADKMAXLLMCALLS environment variable to configure LLM call limits.
  • Create and delete data agent tools added to the dataagent toolset.
  • Model Armor guardrail plugin.
  • Location parameter added to listaccessibledataagents.
  • New samples for NVIDIA NIM integration and skill state injection.

Improved

  • Import performance optimized to reduce startup time.
  • Request path avoids Pydantic round-trips and streaming accumulation inefficiencies.
  • Conversation history handling improved for Gemini round-trip compatibility.
  • Live API now transfers to target agent regardless of function response order.
  • ReadFileTool ranged reads avoid shell execution for better security.
  • Audio stream end event support added for realtime input.
  • Background tool tasks now properly stopped when live agent runs end.
  • Session resumption handle properly used when opening live sessions.

Fixed

  • Shell execution vulnerability in ReadFileTool ranged reads.
  • Tool confirmation security and re-validation issues.
  • Failed inference results handling in eval without invocations.
  • NLTK 3.10.1 dependency issue breaking venvs in working directories.
  • Session initialization event validation.
  • Host header validation on local API server requests.

July 2026

8 releases: 2.4.0 → 2.6.1

July brought significant security and integration improvements to ADK, with new support for mutual TLS authentication, expanded API compatibility, and enhanced agent capabilities. The month saw steady progress across authentication, tooling, and deployment features.

Action needed
  • GCP Skill Registry endpoint changed to AgentRegistry; existing users must update their configuration.

Highlights

  • Added mutual TLS support for Discovery Engine and Google API tools, improving security for external service connections.
  • Introduced support for OpenAI Responses API and asynchronous non-blocking tools in Live API.
  • New GCP Skill Registry integration using AgentRegistry endpoint, plus search agents and MCP server support.
  • Added strict input schema validation for LlmAgent workflow nodes and per-invocation auth headers for agent cards.
  • Expanded Managed Agents API with single-turn mode and improved node input bridging.

New

  • Mutual TLS support for DiscoveryEngineSearchTool and Google API tools.
  • OpenAI Responses API support in labs.
  • Support for reusing existing SQLAlchemy AsyncEngine in DatabaseSessionService.
  • Session TTL and expiration support in Vertex AI session service.
  • Anthropic effort config handling and thinking parameter propagation.
  • ManagedAgent backed by the Managed Agents API.
  • Cloud Run sandbox option in code executors.
  • MCP server capability to serve an ADK agent, plus search agents in agent registry.

Improved

  • Thinking and tool-use token columns now exposed in BigQuery analytics views.
  • GCP Skill Registry upgraded to use agentregistry/skill endpoint.
  • Strict input schema validation added for LlmAgent workflow nodes.
  • ManagedAgent now includes single-turn mode field and improved node input to user content bridging.
  • GenAI SDK support upgraded to 2.9+.
  • Custom Gemini LLM connection logic extended to all 3x models, not just 3.1.
  • ADK deploy agentengine command now supports --extrapackages option.
  • CLI telemetry enhanced with parent terminal grouping, TTL pruning, and TTY connectivity capture.

Fixed

  • Vertex AI Live API session replay on reconnect and modality switch.
  • Connection cleanup on expected GoAway signal in bidi streaming.
  • Tool confirmation validation no longer repeats for already-consumed confirmations.
  • Early telemetry recording for long-running web servers with proper exit code logging.

June 2026

6 releases: 1.34.2 → 2.3.0

June brought stability improvements for Gemini Live models, support for the latest Gemini 3 preview, and new enterprise features including mTLS and enhanced analytics. The month focused on fixing transcription handling and preparing for the upcoming Gemini 2.5 Flash deprecation.

Action needed
  • GenAI SDK v2.0.0 renames turn-based helpers (convertcontentstoturns → convertcontentstosteps) and changes terminology from 'turns' to 'steps'
  • Default agent model changed from gemini-2.5-flash to gemini-3-flash-preview

Highlights

  • Default model for new agents switched to gemini-3-flash-preview, ahead of Gemini 2.5 Flash shutdown in October 2026.
  • Fixed grounding metadata handling in Gemini 3.1 Live that was being silently discarded.
  • Added GCS first-party toolset integration and improved MCP tool error handling to prevent session drops.
  • Introduced AutoTracingPlugin for OpenTelemetry auto-instrumentation and RubricBasedMultiTurnTrajectoryEvaluator.
  • Added mTLS support in AgentRegistry client for enterprise deployments.

New

  • AutoTracingPlugin for OpenTelemetry auto-instrumentation
  • RubricBasedMultiTurnTrajectoryEvaluator for trajectory evaluation
  • GCS first-party toolset integration
  • mTLS support in AgentRegistry client
  • createhttpoptions parameter in ContextCacheConfig for cache creation timeout
  • GEPARootAgentOptimizer for agent optimization
  • Log level option for adk run CLI
  • Request timeout option for loadwebpage tool

Improved

  • Input transcription handling for Gemini Live 3.1 models
  • Grounding metadata now sent only at the end of each turn for Gemini 3.1 Live
  • Streaming ensures final partial=False frame is always yielded
  • BigQuery Agent Analytics reliability improvements
  • API server now aborts runs on client drops to avoid resource leaks
  • Reconnect attempts reset on successful connection
  • User simulator config now exposed in generateresponses for evaluation
  • Core and CLI migrated to enterprise parameters

Fixed

  • Grounding metadata in Gemini 3.1 Live was being silently discarded
  • Input and output transcription finished events for Gemini v3.1
  • MCP tool errors causing session drops
  • Live event buffering in runner removed to improve performance
  • History config injection support added for Gemini 3.1 Live on Vertex AI

May 2026

6 releases: 1.32.0 → 1.34.1

May brought major architectural advances to ADK, including the launch of version 2.0 with flexible execution graphs and multi-agent orchestration capabilities. The month also saw steady improvements to analytics, authentication, and model integrations across the 1.x line.

Highlights

  • Introduced Flexible Execution Graphs enabling non-linear, conditional, and cyclical agent workflows in version 2.0.
  • Added native inter-agent routing for seamless multi-agent collaboration and context propagation.
  • Implemented modular workflow abstractions supporting parallel sub-agent workers and hierarchical team structures.
  • Added native OpenTelemetry agentic metrics and expanded tracing capabilities.
  • Introduced support for Anthropic's thinking blocks and Gemini Live API evaluation.

New

  • Flexible Execution Graphs engine for model-agnostic orchestration of complex agent patterns.
  • Native inter-agent routing with control state handoffs and context variable propagation.
  • Support for Anthropic's thinking blocks in agent responses.
  • Gemini Live API support in ADK evaluate.
  • BufferableSessionService for improved session management.
  • GCPSkillRegistry implementation for skill management.
  • Sandbox creation from templates and snapshots.
  • Chart generation and artifact loading for data agents.

Improved

  • Made ADK environment tools truncation limit configurable.
  • Added credentials parameter to BigQueryAgentAnalyticsPlugin.
  • Express mode onboarding support added to adk deploy CLI.
  • Extended LlmResponse with getfunctioncalls and getfunctionresponses methods.
  • Added option to prevent SaveFilesAsArtifactsPlugin from attaching reference file parts.
  • Support for excluding predefined functions in ComputerUseToolset.
  • Improved fork detection and response logging in BigQuery plugin.
  • Added config option to include tool calls and responses in conversation history for user simulator.

Fixed

  • Fixed double append bug affecting session handling.
  • Corrected grounding metadata handling in Gemini 3.1 live that was being silently discarded.
  • Fixed input and output transcription finished events for Gemini v3.1.
  • Improved error handling when sandbox is missing by catching genai.ClientError.
  • Fixed session drop on MCP tool errors.
  • Fixed video events with inline data being stored in session.

April 2026

5 releases: 1.28.1 → 1.31.1

April brought significant enhancements to Agent Development Kit's debugging and testing capabilities, including a redesigned Live UI with new visualization tools and improved session management. The month also expanded model support, added security improvements, and introduced new integrations for authentication and resource management.

Action needed
  • Arguments are now disallowed on /builder endpoint.

Highlights

  • Completely redesigned Live UI with new trace view, graph view, and computer use visualization for better agent debugging.
  • Added support for Gemma 4 models and improved live preview capabilities with the gemini-3.1-flash-live-preview model.
  • Introduced session display names, event filtering, and structured execution logging for easier agent monitoring.
  • Enhanced security with shell metacharacter blocking in BashTool and fixes for YAML configuration vulnerabilities.
  • Expanded integrations with Auth Provider support in agent registry, Parameter Manager, and live avatar support.

New

  • Live UI with redesigned real-time chat and testing interface.
  • Graph view for visual mapping of agent architectures and node relationships.
  • Computer use view for visualizing screenshots and logs during computer-use tasks.
  • Session display names for organizing and tracking debug sessions.
  • Event filtering by message type, tools, or errors.
  • Auth Provider support in agent registry with credential management for MCP toolsets.
  • Parameter Manager integration to ADK.
  • Gemma 4 model support in ADK.

Improved

  • Live tool call handling now buffers and emits tool calls together upon turn completion.
  • BashTool now supports configurable resource limits for subprocesses.
  • BigQueryLoggerConfig now has configurable viewprefix option.
  • Custom session ID functionality added to Vertex AI session service.
  • Regional endpoint support added to SecretManagerClient.
  • Live session resumption updates now exposed as events in BaseLlmFlow.
  • BigQuery tools promoted to Stable status.
  • Live avatar support added to ADK.

Fixed

  • Security vulnerability patched for remote code execution via nested YAML configurations.
  • TaskStatusUpdateEvent now emitted for ADK events with actions but no output parts.
  • Web OAuth flow and trace view functionality restored.
  • Bound token for MCPTool disabled to improve security.

March 2026

7 releases: 1.27.0 → 1.28.0

March brought improvements to agent-to-agent communication, expanded LLM model support, and enhanced authentication capabilities. The month also included several security and stability updates across multiple releases.

Highlights

  • Agent-to-agent communication now supports request interceptors and configurable lifespan parameters
  • Anthropic LLM models now support PDF documents and streaming responses
  • New pluggable authentication system allows custom auth provider integrations
  • Added MultiTurn Task metrics for tracking success, trajectory, and tool usage
  • Slack integration is now available for ADK

New

  • A2A request interceptors in RemoteA2aAgent
  • UiWidget support in EventActions for experimental UI Widgets feature
  • Pluggable AuthProviderRegistry for custom authentication integrations
  • PDF document support for Anthropic LLM
  • Streaming support for Anthropic models
  • Durable runtime support
  • Fine-grained access control via database role property in SpannerToolSettings
  • MultiTurn Task metrics including success, trajectory, and tool trajectory tracking

Improved

  • LLM Agent now supports all types.SchemaUnion as output schema
  • GetSessionConfig now passed through Runner to session service
  • BigQuery Toolset OAuth scope configuration improved
  • Vertex AI usage metadata storage and retrieval via custom metadata
  • EventActions expanded to support new experimental UI Widgets
  • Events table performance with added index
  • Individual unit tests can now be run via unittests.sh
  • Slack integration added to ADK

Fixed

  • Fixed issue affecting LlmAgent creation due to missing version field
  • Fixed BigQuery Toolset OAuth scope for Dataplex
  • Added protection against arbitrary module imports
  • Excluded compromised LiteLLM versions from dependencies
  • Gated builder endpoints behind web flag for security

February 2026

5 releases: 1.24.0 → 1.26.0

February brought significant improvements to the developer experience with enhanced debugging tools, accessibility features, and expanded LLM integrations. The month focused on streamlining workflows and adding support for new deployment and optimization capabilities.

Action needed
  • Credential manager now accepts toolcontext instead of callcontext.

Highlights

  • Redesigned event debugging with click-to-expand message rows for faster troubleshooting.
  • Full keyboard navigation support with arrow keys for improved accessibility.
  • Integrated A2UI v0.8 standard catalog to automatically render UI components in chat.
  • Added token compaction during and before requests to optimize context management.
  • Expanded LLM support with Apigee streaming integration and LiteLLM model registry.

New

  • Agent optimizer interface for building custom optimization workflows.
  • Demo for simple prompt optimizer showing optimization interface capabilities.
  • Autocreate session flag for ADK API server to streamline initialization.
  • Post-invocation token-threshold compaction with configurable event retention.
  • Conformance test report generation for validation and compliance tracking.
  • Load skill from directory method for easier skill management.
  • Agent Skills spec compliance with validation, aliases, scripts, and auto-injection.
  • BigQuery ADK support for search catalog tool integration.

Improved

  • Detailed tooltips for function calls showing arguments, responses, and state changes.
  • Event ID now passed as metadata when converted into messages.
  • Agent simulator improved with better prompts and environment data integration.
  • Google credentials configuration now accepts externally passed access tokens.
  • Gemini LLM class now supports custom base URL option.
  • Apigee LLM extended with /chat/completions endpoint integration and streaming support.
  • Intra-invocation compaction added alongside pre-request token compaction.
  • Memory service URI can now be specified in ADK CLI run command.

Fixed

  • Fixed pickling lock errors in McpSessionManager.
  • Fixed A2UI component types.
  • Restored deprecated eval endpoint for web migrations.

January 2026

3 releases: 1.22.0 → 1.23.0

January brought expanded agent capabilities, improved observability, and better developer tooling. The release cycle focused on optional configurations, simulation enhancements, and standardized tracing.

Action needed
  • BigQuery plugin tracing now uses OpenTelemetry instead of custom ContextVar implementation; existing tracing configurations must be updated.

Highlights

  • Progressive server-sent streaming is now enabled by default for faster response handling.
  • BigQuery plugin now uses OpenTelemetry for tracing, replacing the previous custom implementation.
  • Sessions can now be automatically created when needed, reducing setup overhead.
  • Custom instructions support added to user simulators for more realistic testing scenarios.
  • New Cloud Pub/Sub integration tool available for agents.

New

  • Custom instructions support for LlmBackedUserSimulator to enhance simulation realism.
  • Per-turn evaluator for user simulations to assess agent performance within conversations.
  • Handwritten tool for Cloud Pub/Sub messaging integration.
  • Automatic session creation when a session does not exist.
  • Support for token endpoint authentication methods in OAuth2 credentials.
  • Computer use field added to agent information in adk-web server.
  • Thinking config support in generatecontentconfig.
  • CLI option to disable features and override default feature enable state.

Improved

  • LlmAgent.model is now optional with automatic fallback to a default model.
  • Regex patterns now supported for allowed origins configuration.
  • Custom httpx factory parameter exposed for MCPS streamable HTTP customization.
  • A2UI message conversion between A2A DataPart metadata and ADK events.
  • OpenTelemetry integration added to adk deploy agentengine command with oteltocloud flag.
  • AgentEngineSandboxCodeExecutor removed from experimental status, marking it as stable.

Fixed

  • adk migrate session CLI command restored.
  • Database reserved keyword escaping applied.

December 2025

2 releases: 1.20.0 → 1.21.0

December brought expanded LLM provider support, enhanced streaming capabilities, and new integrations for agent development. The month focused on enabling more flexible agent configurations and adding support for emerging AI model features.

Highlights

  • Multi-provider support now available for Claude and LiteLLM alongside existing options.
  • Google Gemini Interactions API is now supported in ADK.
  • Streaming function call arguments enabled through progressive SSE streaming.
  • APIRegistryToolset lets you add tools from Cloud API registry directly to agents.
  • BigQueryAgentAnalyticsPlugin upgraded to v2.0 with improved performance and multimodal support.

New

  • APIRegistryToolset for integrating Cloud API registry tools into agents.
  • Gemini Interactions API support.
  • Image parts support in user messages for Anthropic Claude.
  • Endpoint to list apps with details.
  • addSessionToMemory method in CallbackContext and ToolContext for explicit session persistence.
  • ADK EventActions included in agent-to-agent responses.

Improved

  • BigQueryAgentAnalyticsPlugin v2.0 with performance enhancements and multimodal support.
  • Streaming function call arguments now handled in progressive SSE streaming.
  • BigQueryAgentAnalytics plugin now includes location information for tables in agent events.
  • Multi-provider support expanded to include Claude and LiteLLM options.

Fixed

  • Validation added for unique sub-agent names to prevent configuration errors.
  • Option added to prevent propagation of runner plugins to AgentTool runners.

Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. Agent Development Kit (ADK): pricing, features and alternatives · All changelogs