Repository navigation
2.15.0 (2026-11-19)
Highlights
This release keeps voice streams and agent turns responsive under slow tools and session storage, measures multi-run reliability and partial tool-call accuracy, isolates sandboxes per user or session, adds Google Maps and project-wide BigQuery table discovery, and speeds up RAG memory search up to 6x. It also removes the interactive Visual Agent Builder from ADK Web.
- Integrations:
- Verify addresses, reverse-geocode coordinates, and fetch Street View imagery directly from an agent with
GoogleMapsToolset(get_street_view_image,validate_address, andreverse_geocode). See the unit guide and sample for more details. (5371bb1) - Answer open-ended analytics questions without knowing exact
table_referencesupfront:BigQueryToolset'sask_data_insightsnow acceptssearch_scope(includeProjectIdsandincludeDatasetReferences) to discover relevant tables across projects or datasets. (9eb6d9d)
- Verify addresses, reverse-geocode coordinates, and fetch Street View imagery directly from an agent with
- Sessions and artifacts:
- Keep slow database or network writes off the critical path of agent turns with experimental
BufferedSessionServiceandBaseSessionService.append_events, which buffer session events in memory and flush them in background batches. See the unit guide for more details. (e1deb17, f4fab78) - Use a single artifact service across local development and cloud object stores (
gs://,s3://,file://, andmemory://):FsspecArtifactServicenow writes, deletes, and versions artifacts (save_artifact,delete_artifact,get_artifact_version, andlist_artifact_versions), including S3 custom metadata. (df9f9d0, b916291)
- Keep slow database or network writes off the critical path of agent turns with experimental
- Evaluation:
- Distinguish whether an agent can solve a task at least once (
pass_at_k/pass@k) versus reliably on every run (pass_hat_k/pass^k) inmulti_run_config:adk evalnow reports their mean across the eval set, prints a per-case multi-run summary table by default, and savesMultiRunCaseSummaryresults to.evalset_result.json. See the unit guide for more details. (5c7a829, b64b540, 287c0b0) - Score tool calls with partial credit per argument instead of failing the whole call on one mismatched parameter using
tool_parameter_match(ToolParameterMatchCriterion), backed byToolArgsMatcherandIN_ORDER,EXACT, orANY_ORDERcall pairing. See the unit guide for more details. (04ba412)
- Distinguish whether an agent can solve a task at least once (
- Execution environments: Isolate sandboxes per user or session without making other concurrent sessions wait during sandbox startup:
BaseEnvironmentmethods now receive the calling run'sctx: Context, andBaseEnvironment.scope(EnvironmentScope.SHARED,USER, orSESSION) scopes initialization locks accordingly. See the unit guide for more details. (1bd22e3, 691ecbe) - Live streaming and performance:
- Keep voice audio, transcriptions, and barge-ins flowing while slow tools execute:
run_livenow runs every tool call in a background task, andBaseLlmConnection.send_context_updateupdates system instructions or tools mid-session on open Gemini 3.x Live connections without reconnecting. See the unit guide for more details. (e5dd820, 6ccd07b) - Cut RAG memory search latency up to 6x in
VertexAiRagMemoryServiceandVertexAiRagRetrievalby reusing API clients per event loop, and start parallel fan-out branches concurrently in resumableWorkflowruns by coalescing status checkpoints once per scheduler tick. (112491d, 6dbe5bb)
- Keep voice audio, transcriptions, and barge-ins flowing while slow tools execute:
Breaking changes
- Interactive Visual Agent Builder removed from ADK Web: Developers increasingly build agents by co-building with a coding agent rather than wiring nodes by hand, and builder usage has declined steadily. ADK Web now focuses on running, testing, debugging, and evaluating agents. Agents you built with the builder are standard Agent Config YAML and still run with
adk web,adk run, andadk deploy. To create new agents, use a coding agent such as Antigravity with the Agents CLI skills, or start fromadk create. The/dev/apps/{app_name}/builder*routes andgoogle.adk.cli.built_in_agentsare removed. (c890aa3, e2a21ac, fb08c78) - Explicit prompt arguments in
McpInstructionProvider(security hardening): To avoid exposing unrelatedsession.statekeys to external MCP servers,McpInstructionProviderno longer auto-fills prompt arguments fromsession.state. PassMcpInstructionProvider(..., arguments=...)(a mapping or a callable overReadonlyContext) for any arguments the prompt expects. (889ddd3)
Behavior changes
- Live function calls run in background tasks:
run_liveexecutes all tool calls in background tasks without blocking audio or interruptions; parallel calls emit separateFunctionResponseevents as each finishes, and tool exceptions return an error response instead of ending the live run. (e5dd820) DYNAMIC_INSTRUCTION_ROUTINGenabled by default: Dynamic instructions fromload_artifactsandload_mcp_resourceare placed in a user-role message before the current turn instead of appended tosystem_instructionso the static instruction prefix stays cacheable (setADK_DISABLE_DYNAMIC_INSTRUCTION_ROUTING=1to opt out). (c6586f1)
All changes
Features
- add BaseLlmConnection.send_context_update (6ccd07b)
- add CheckGroundingClient for grounding verification (abcc020)
- add samples for context filter, Model Armor and BigQuery plugins (389f80a)
- artifacts: add write, delete, and version metadata support to FsspecArtifactService (df9f9d0)
- artifacts: support artifact metadata on S3 in FsspecArtifactService (b916291)
- attach agent cached_content seed on initial turn (3256db0), closes #7536
- bigquery: add search_scope support to ask_data_insights (9eb6d9d)
- environment: add
EnvironmentScopeand lockinitializeper scope (691ecbe) - environment: pass the calling context to BaseEnvironment methods (1bd22e3)
- eval: add pass@k and pass^k as multi-run cross-run criteria (5c7a829)
- eval: persist multi-run case summaries and print default Pass Rate summary (287c0b0)
- eval: report each cross-run criterion's mean over the eval set (b64b540)
- evaluation: add configurable tool parameter match metric (04ba412), closes #6726
- features: enable DYNAMIC_INSTRUCTION_ROUTING by default (c6586f1)
- integrations: add Google Maps Platform integration (5371bb1)
- let toolsets opt out of caching their tools per invocation (b979abe)
- list bundled resource paths in the load_skill response (c41bc42)
- live: add client label telemetry for CascadeLive reasoning turns (9620b42)
- live: run every live function call in a background task (e5dd820)
- models: add LlmRequest.to_generate_content_body() (9be8630)
- plugins: add eval session and turn query templates with filters (5e83dd5)
- plugins: isolate nested AgentTool state and add response_text_key (5ce1d0c)
- remove the Agent Builder Assistant and built-in
__agents (e2a21ac) - remove the Agent Builder endpoints from the dev server (c890aa3)
- sessions: add BaseSessionService.append_events and batch buffered writes with it (f4fab78)
- sessions: add BufferedSessionService to write session events in the background (e1deb17)
- support bounded parallel agent workers (5e4f0e1), closes #7641
- telemetry: record the inference token metrics in every schema version (1de9760)
- telemetry: tag Cloud Run and Kubernetes runtimes in client labels (1997a12)
- telemetry: use the run's telemetry config for model_consult advisor metrics (f9d948c)
Bug Fixes
- a2a: persist streamed RemoteA2aAgent output without last_chunk (dbae28d), closes #7255
- a2a: resolve finish_task args from unary A2A task history (1cb926d), closes #7246
- accept AgentTool input_schema args that use aliases (90d352b)
- accept content lists in Interactions code execution results (6959bac)
- accept every SDK thinking level in ModelConsultTool (c74df83)
- antigravity: pass cached token count in the turn usage event (cd8debd)
- build: set uv exclude-newer to 7 days in ADK pyproject.toml (f390bd9)
- capture app_details debug log on the google_adk logger in its test (b4bf854)
- clean up empty maps after deleting sessions (c0bcdaa), closes #7722
- cli: keep extra_packages before requirements and reinstall after (5e91ab0)
- cli: log the redacted payload /run returns (ba774bc)
- cli: require a local client for the telemetry consent write (3e6c586)
- cli: return a fixed detail string when saving an artifact fails (6507ad8)
- cli: validate agent import on Agent Engine container startup (7d2c7b7), closes #8062
- deduplicate SlackRunner execution on thread mentions (d103e36)
- do not promote a failed remote A2A task to workflow output (f67cc35)
- do not score hallucinations_v1 responses with unlabelled sentences (59b0dd6)
- drop parent_agent from BaseAgent's repr (8ec642c)
- emit no empty event from BuiltInCodeExecutor (61f8143), closes #7878
- end every Claude request on a user turn (92b93c8), closes #7861
- enforce ExecuteBashTool resource limits that are set to 0 (a702e3c), closes #7983
- eval: keep the session details of every eval case in its result (f1b1f9d)
- eval: send Gemini 4 a thinking level instead of the default budget (36ac021)
- eval: stop judge and Vertex metrics from scoring model thoughts (dd7161f)
- exclude ADK client function calls from evaluated tool calls (825a374), closes #8025
- fail on connector schema operation error and bound poll wait (ab63844), closes #7165
- flush debug logging trace and clear state on unhandled run errors (9e549ad)
- give a cloned LlmAgent tools for its own sub-agents (61a4dc6), closes #7894
- honor parallelism_limit in simulator and hallucination evals (656696c)
- honor subclass validation aliases in Event and LlmAgent kwargs (72663e5)
- initialize the environment from tool calls (61efa30)
- keep $defs of a recursive output_schema in OpenAI strict mode (245f5d5)
- keep a delegation pending while its workflow tool awaits input (6012810), closes #7974
- keep a text answer as text when the input request asks for a string (b989d68)
- keep Chinese and Japanese lines in load_web_page (89ae4d3), closes #7956
- keep daytona on the 0.x SDK DaytonaEnvironment uses (4da8071)
- keep ManagedAgent Developer API clients on v1beta (30cb66e)
- keep streamed OCI parallel tool calls separate and in order (353a771)
- keep the Gemini API on v1beta unless a version is configured (aa0f434)
- keep the live resumption handle through non-resumable updates (02db0ff)
- leave paused long-running calls open when the run task is cancelled (e1e231c)
- let MCP request headers replace static headers in any case (58d1ebe)
- live: send tracking headers on live sessions through caller-supplied clients (761538a)
- live: yield finished tool responses when the live connection fails (1a0ddab)
- live: yield interrupted before saving the model's audio on a barge-in (907414b)
- mask credential material in the telemetry serializer (b695287), closes #7311
- mcp: retry a tool call once after the server drops the session (10711d0), closes #6822 #6933
- name the calling node in the rerun_on_resume error (ba43b4f)
- only create the agents file watcher when agent reload is enabled (12aab4a)
- only pip install extra_packages folders that declare a project (c16dac7)
- pass a single_turn transfer target's reply downstream (b4126cd), closes #7782
- pass each adk deploy --allow_origins as its own server argument (1ade0ba)
- plugins: answer failed tool calls in eval traces and skip HITL tools (10b7162)
- plugins: close wrapped generators in AutoTracingPlugin (0192edd)
- plugins: mark truncated eval turns not ready (16898f3)
- preserve empty Anthropic thinking blocks (946ca77)
- preserve line endings in FileArtifactService text artifacts (1fa0044), closes #7914
- preserve required flag in RestApiTool serialization (e5ffe9c), closes #8065
- raise the SqliteSessionService busy timeout to 30 seconds (d09599e)
- read LangGraphAgent reply text from content blocks (0795861)
- repair the model samples that fail on their first turn (13b6ae8)
- report empty Claude responses as model errors (5839007), closes #7660
- require confirmation before large reimbursements in hitl samples (982f608)
- respect display_name from .agent_engine_config.json (319b5a7), closes #7619
- run a node once when a list of routes matches several edges to it (dc7d743), closes #8072
- run a task agent's after callbacks when finish_task succeeds (fa647fc)
- run code without an artifact service when it writes no files (9687bae)
- run OAuth2 auth response token exchange off the event loop (f8ba96a)
- run service account token exchange off the event loop (9651df0)
- run synchronous metric evaluators off the event loop (67b217d)
- score each invocation against its own rubrics in rubric-based metrics (e4f6c3a)
- send gs:// file parts to Claude as inline base64 (e87ef73)
- send max_completion_tokens for Apigee reasoning models (32c202e), closes #7863
- send media a tool returns to the Interactions API (80bd338)
- send structured output schemas on Apigee chat completions (e648b71), closes #7864
- sessions: isolate stored events and state deltas in append_event (a0339b7), closes #7701
- show the working directory based on
working_dirinstead ofis_initialized(d05e482) - skills: refresh a stale skill copy in the environment when its script runs (d94b5e2)
- skills: stop writing skills into the environment while a request is built (5f2f227)
- skip the genai OTel instrumentor when dependencies conflict (277553a)
- split mixed-script search queries in InMemoryMemoryService (e9f9790)
- stop concurrent writes losing app: and user: state (6b2d455)
- stop emitting
required: trueinside optimization field schemas (3bb410c) - stop get_session dropping events appended out of timestamp order (b7c9c40)
- stop MCP servers from reading session state through prompt arguments (889ddd3)
- stop rerunning a waiting wait_for_output node on resume (ba142cf), closes #7905
- stop retrying context cache creation on models that refuse it (43e959e)
- stop returning uncaught exception text to remote A2A peers and trigger callers (d61f428)
- stop tagging YAML agents as Visual Builder apps in telemetry (fb08c78)
- stop the multimodal plugin resending inline media in session mode (2b7e320)
- support google-genai releases that move converters to _converters (15af24c)
- support non-LLM root nodes in app-info and get_agents_dict (ec2ac7c)
- take the first non-null union member in _remove_any_of (ca42dcd)
- telemetry: exclude tool runtime from model call duration without a recorded span (8d857f4)
- telemetry: send the hosting runtime tag only to Google API endpoints (c38ecda)
- telemetry: set gen_ai.response.model on LLM spans (ecc6b6e), closes #7243
- telemetry: use semconv units for invoke_agent call-count metrics (56b25c6)
- tools: close child runner in a finally block in AgentTool.run_async (73cfd62)
- tools: report the thinking level ModelConsultTool actually sent (8fa57e8)
- trace sub-agents by the model they inherit (5112046), closes #7892
- treat Gemini 4 and later as Gemini 3 or newer in LiteLlm (de1d307)
- type agent loaders as returning any workflow node (f0086e3)
- type ComputerUseToolset's pickled state (ddb812e)
- type the API server eval managers as optional (bd43897)
- upgrade watchdog for Python 3.14 macOS wheels (7941f8d)
- use a thinking level for SemanticArgMatcher on Gemini 4 (fa7b204)
- use agentidentitycredentials v1 client in /agent-identity/finalize (f0eb613)
- use standard tool role for Gemma 4 on Vertex AI (b33b43e), closes #7977
- vmaas: AgentEngineEnvironment can be pickled after use (fa9ae7e)
- workflow: clear open interrupts sealed by error function responses (efc7717)
- workflow: ignore a resume response schema of an unsupported shape (01b4b5a)
- workflow: isolate and preserve user_content for single-turn workflow nodes (1756822), closes #7944
- workflow: key a resumed sub-agent's run by the path its events carry (75fc853)
Performance Improvements
- reuse RAG API clients per event loop, cutting memory search 6x (112491d)
- workflow: coalesce node status checkpoints per scheduler tick (6dbe5bb)
Documentation
- cli: document multimodal Content dict inputs for Agent Engine queries (11d088b)
- correct agent skill claims the code no longer supports (dd14fa2)
- fix syntax errors and language tags in sample READMEs (68ee084), closes #7998
- plugins: add a unit guide for the eval query builders (278fe05)
- remove experimental and allowlist notes for deferred service tier (1973796)
- samples: add an OpenAI Responses API sample with a reasoning model (ef812bd)
- say tool_filter does not apply to load_mcp_resource (b79c529)
- use the current Agent Runtime product name in README (e41b46b)