Google ADK Python v2.6.0 adds features, fixes
sdkaiengineermediaenergypython-agentic
feature patch
Google's ADK Python library has been updated to version 2.6.0, introducing numerous new features and improvements. Key additions include enhanced support for authentication, new evaluation samples for agents, expanded tooling for Eventarc and BigQuery, and improved telemetry and CLI capabilities. These changes aim to provide developers with more robust tools for building and deploying AI agents, with a particular focus on enhanced flexibility and integration within the Google ecosystem.
- →support per-invocation auth headers when fetching agent cards
- →add --extra_packages option to adk deploy agent_engine , closes #3936
- →Add ADK evaluation samples: shared home-automation agent
- →add agent identity auth manager finalize endpoint for 3 legged OAuth flow with auth manager
- →Add basic_criteria ADK eval sample
Features (37) ›
- a2a support per-invocation auth headers when fetching agent cards
- add --extra_packages option to adk deploy agent_engine , closes #3936
- Add ADK evaluation samples: shared home-automation agent
- add agent identity auth manager finalize endpoint for 3 legged OAuth flow with auth manager
- Add basic_criteria ADK eval sample
- Add bigquery-graph skill to ADK BigQuery tools
- add capability to log commands run in CLI
- Add custom_metric ADK eval sample
- Add llm_judge_match ADK eval sample
- add metadata extraction symmetry to A2A converters
- add opt-in final_response_tool_names to BigQueryAgentAnalyticsPlugin
- Add ReflectAndRetryModelPlugin for self-healing model errors
- Add rubric_criteria ADK eval sample
- add state_delta support to LiveRequest for live mode , closes #4220
- add telemetry consent check, status commands, and interrupt safety to CLI
- Add telemetry consent configuration endpoints and local writing utility
- add telemetry metrics collection for ADK CLI execution
- Add test_file_vs_evalset ADK eval sample
- Add user_simulation ADK eval sample
- agents add instruction field to ManagedAgent
- agents forward ManagedAgent instruction as system_instruction
- eval Make live and audio evals reachable via public entrypoints
- eventarc add Eventarc Advanced toolset for ADK
- eventarc add Eventarc Advanced toolset for ADK
- forward the OAuth2 nonce to the authorization request , closes #2067
- integrations add OCI Generative AI provider , closes #5069
- make agent evaluation compatible with pre-loaded artifacts , closes #2075
- make EventsCompactionConfig sliding-window fields optional , closes #6398
Publish companion constraints-3.11.txt and constraints-3.12.txt file for transitive dependency protection (4 day buffer to protect from supply chain attack)
- refresh expired OAuth2 tokens in OpenAPI credential exchanger
- samples add ManagedAgent create-and-use custom-agent sample
- support audio generation user simulator
- support serving a to_a2a agent under a path prefix , closes #4448
- support user labels in RunConfig
- telemetry add request-driven metric export for Agent Engine
- telemetry derive gen_ai error.type from provider status code
- Update agent_registry to handle mTLS endpoints internally
Enhancements (4) ›
- avoid quadratic text/audio accumulation in streaming
- cache the FunctionTool declaration across LLM calls
- import lazy-load heavy dependencies
- tools avoid double Schema serialization in Optional/Union dedup
Fixes (82) ›
- a2a honor task cancellation instead of raising NotImplementedError
- agents await cancelled tasks in pre-3.11 ParallelAgent merge , closes #5297
- agents skip output_key processing on intermediate conversational text in task-mode LlmAgent
- allow http_options.extra_body in generate_content_config
- allow invocation-level rubrics
- artifacts namespace file artifacts by app
- assemble only the current turn
- bound integration HTTP waits
- bypass Protobuf Gencode/Runtime version check in antigravity integration
- caching prevent prompt cache invalidation when using dynamic tools , closes #3227
- Call mtls.should_use_mtls_endpoint with client_cert_available argument
- canonicalize context cache fingerprint for stable hashing
- ci mark imported PRs as merged even if already closed
- cli treat agent folder with subfolders as single agent in adk web
- close A2A response stream when the caller stops consuming
- close AsyncDaytona client in DaytonaEnvironment.close
- collect all tools so native tools aren't dropped , closes #6091
- derive an SDK-conforming Antigravity conversation id
- do not mount a cluster credential into the GKE code sandbox
- emit only new artifact parts on streaming artifact updates , closes #6343
- emit resumability checkpoints from workflow graph nodes
- eval support get_agent_async in adk eval
- evaluation support non-English responses in ROUGE-1 matching
- guard multimodal tool results plugin against empty contents
- handle Windows paths in adk eval , closes #6415
- harden A2A metadata serialization and parser type validation
- include credential failure details
- isolate delegated task branches , closes #6457
- let httpx set the multipart boundary Content-Type
- make InvocationEvent.content optional so the Web UI can save eval cases , closes #6336
- make ParallelWorker concurrent failure exceptions deterministic
- match rubric verdicts by echoed id so paraphrased rubrics are not dropped , closes #6171
- memory honor Vertex RAG top-k configuration
- memory make Vertex RAG uploads async-safe
- migrate v0 event timestamps as local time, not UTC
- models capture Anthropic thinking-block signatures during streaming
- models guard Gemini.client_kwargs against missing-field AttributeError
- models populate finish_reason on AnthropicLlm responses
- narrow broad except in agent identity credential providers
- percent-encode context id fields so ids with separators round-trip
- pin third-party github actions to commit shas
- plugins complete BigQuery Agent Analytics privacy and shutdown hardening
- plugins correct lazy import path for ReflectAndRetryModelPlugin
- plugins harden BigQuery agent analytics against fail-open privacy, GCS concurrency, and startup-loss gaps
- plugins restore failure counter properties on retry tool plugin
- populate finish_reason on Anthropic LLM responses
- populate function name in FunctionResponse parts
- present client certificate and use mTLS endpoint for API Hub calls
- preserve ADK behavior on Windows
- preserve explicit false and zero OpenAPI query parameters , closes #6287
- preserve metadata in A2A artifact updates
- Prevent adk deploy from uploading .adk/session.db file
- Prevent continuation forgery in tool confirmation
- prevent transfer_to_agent loop on resumable invocation replay
- propagate custom metadata from RunConfig to InvocationContext
- raise SessionNotFoundError when appending to a missing session
- register custom metrics from eval config in LocalEvalSampler , closes #6177
- reject base_url and extra_body in generate_content_config
- reject incomplete evaluation inputs
- require patched async LangGraph runtime
- resolve tool confirmation resumption failure in production
- respect A2A Message.role in inbound event conversion , closes #5186
- return empty list from _get_required_fields when no properties , closes #5920
- scope replay sequence to the current invocation , closes #6497
- scope the tool thread pool to its event loop
- send the bare input_schema payload from AgentTool again
- serialize eval criteria as their concrete subclass
- serialize raw bytes as base64 in A2A converters
- sessions apply after_timestamp and num_recent_events together in VertexAiSessionService
- share one event id across the partial chunks of a streaming response , closes #1006
- single-flight Discovery Engine mode detection , closes #6101
- stop automatic function calling from failing on Vertex AI return types , closes #3543
- stop tracing credentials passed via config.http_options
- strip markdown code fences before validating output_schema JSON
- support nested agent paths in dot_adk_folder resolution
- surface sub-agent RPC errors from AgentTool
- tolerate a malformed traceparent header in the span processor
- tools bind JSONDecodeError in ToolConnectionAnalyzer.analyze
- treat GitHub content as untrusted in the adk_team sample agents
- wait for in-flight BigQuery writes
- workflow fix task agent resumption in nested workflows
- wrap input_schema payload in ReAct prompt and propagate tool_choice to LiteLLM
Maintenance (6) ›
- Add developer unit guide for ReflectAndRetryToolPlugin
- agents add system-instruction sample and guide section
- document creating and using a custom managed agent
- document workflow resumability model and direction
- explain how an accepted pull request lands
- fix broken relative links in documentation
Read the original announcement →
https://github.com/google/adk-python/releases/tag/v2.6.0
Related releases
- crewAI 1.15.10 Patch Release CrewAI Releases ·
- LangGraph checkpoint-postgres 3.1.1 update LangGraph Releases ·
- LangGraph checkpoint-sqlite 3.1.1 released LangGraph Releases ·
- adk-python adds async non-blocking tool support Google ADK (Python) Releases ·
- crewAI 1.15.9: Improved Tool Failure Reporting and Flow Events CrewAI Releases ·
- pydantic-ai v2.21.0 adds token limit and refreshes model names Pydantic AI Releases ·