google/adk-go v2.7.0

on GitHub

Highlights

v2.7.0 adds Anthropic Claude as a first-party model, search-gated tool discovery for large tool catalogs, an experimental classifier API for workflow routing, and a per-invocation budget on model calls. It also switches telemetry to the OpenTelemetry GenAI semantic conventions by default, which is a breaking change for anyone consuming ADK traces.

• Anthropic Claude models (experimental): model/anthropicmodel implements model.LLM over the Anthropic Messages API and Vertex AI, with streaming, tool calling, image and PDF input, extended and adaptive thinking, structured output and prompt caching. See examples/anthropic. (#1866)
• Search-gated tool discovery: tool/toolsearch wraps any tool.Toolset so the model starts with a search_tools tool plus a configured core set, and the tools it finds are declared from the next model step on. This keeps a large MCP catalog out of every request. See examples/toolsearch. (#1734, #1735)
• Classifiers (experimental): exp/classifier defines a Classifier that answers typed Choice, Binary and Score questions with probabilities. It ships with a client for TypeSafe's Jev decision models (exp/classifier/typesafe), an LLMClassifier backed by any model.LLM, and a workflow Node that routes on the answer, matching adk-python's ClassifierNode. (#2003, #2005, #2035)
• Model-call budget: RunConfig.MaxLLMCalls bounds the model calls one invocation may make, 500 by default, and ends the run with agent.ErrLLMCallsLimitExceeded when the budget is spent. Matches adk-python's max_llm_calls. This changes behavior for existing users, see below. (#1412)
• Telemetry: traces follow OpenTelemetry GenAI semantic conventions v2 by default, as adk-python does. This is breaking for telemetry consumers, see below. (#1635)
• OpenAI-compatible models: images and files in InlineData and FileData parts are sent on both the Responses and Chat Completions endpoints (#1543). A response schema title that is not a valid format name falls back to adk_response instead of a 400 from the API (#1912). A GenerateContentConfig field that a newer genai adds is refused with ErrUnsupportedConfigField instead of being dropped silently (#1922).
• Gemini built-in tools: geminitool.GoogleMaps, geminitool.UrlContext and geminitool.EnterpriseWebSearch. (#2026)
• MCP: with Config.Auth set, each user gets their own MCP session, so servers that bind a session to the user who opened it stop answering other users with 403 (#1755). MCP HTTP requests identify ADK in User-Agent and X-Goog-Api-Client, as model calls already do (#1862).
• Auth fix: credentials from auth.ADC() and auth.ServiceAccount() failed every token mint with context canceled on local gcloud ADC and impersonated service accounts. Services on GCE or with a key file were not affected. (#1904)
• Tool auth groundwork: tool/toolauth adds adk-python's auth data model and wire format for adk_request_credential, so requests, replies and credential keys agree with adk-python and ADK Web. Nothing uses it yet. (#2032)
• Reliability:
  • Workflow joins resume with completed predecessors restored instead of rerunning them. (#1572)
  • LoopAgent stops on context cancellation. WithAgentTimeout and WithAgentCancel work on tool and callback contexts instead of returning nil. (#1810, #1604)
  • Artifacts saved in agent and model callbacks are recorded in ArtifactDelta. OutputKey is saved when a before-agent callback short-circuits. (#1878, #1800, #1963)
  • Optional instruction placeholders such as {key?} and {artifact.name?} surface infrastructure errors instead of rendering an empty string. This changes behavior, see below. (#1748)
  • Vertex AI sessions persist live transcriptions. In-memory memory search skips thought parts. (#1964, #1975)
  • Streaming function tools preserve consumer panics. agenttool keeps code-execution output in sub-agent results. Running an agent without Run returns an error instead of doing nothing. (#1961, #1860, #1850)
  • launcher.Config.TelemetryOptions honors WithOtelToCloud. Live reconnect matches close code 1008 by type, not by error text. (#1968, #1993)
• CLI: adkgo deploy agentengine --mem_deploy no longer defaults Memory Bank to gemini-2.5-flash, which Vertex AI stopped serving to new projects. Memory Bank applies its own default model unless --mem_model is set. (#1902)
• Agent Analytics plugin (plugin/agentanalytics, its own module and tag): aligned with adk-python's schema and event vocabulary, with rich session, HITL, A2A and workflow events, GCS offloading of multimodal content, opt-in exactly-once delivery, drop statistics, hierarchical span ids with latency and time-to-first-token, and 25 pre-built per-event BigQuery views. (#1786, #1787, #1788, #1789, #1790, #1791 and follow-up fixes)

Breaking changes

• Telemetry schema: ADK now emits the OpenTelemetry GenAI semantic conventions v2. The legacy log events are replaced by one gen_ai.client.inference.operation.details event per model call, message content on spans is a structured value rather than a JSON string, execute_tool records arguments and results only when span content capture is on (under gen_ai.tool.call.*), and an LLM-agent root gets a new invoke_workflow parent span. Set ADK_TELEMETRY_SCHEMA_VERSION_OPT_IN=otel_semconv_1_36 to restore the previous format, supported until at least March 31, 2027. (#1635)
• Model calls are capped at 500 per invocation: a run that makes more than 500 model calls now ends with an error wrapping agent.ErrLLMCallsLimitExceeded. Raise the cap with RunConfig.MaxLLMCalls or ADK_MAX_LLM_CALLS, or set MaxLLMCalls to -1 for no limit. Unlike adk-python, 0 means the default, not unlimited. (#1412)
• Optional instruction placeholders propagate errors: {key?} and {artifact.name?} render empty only when the key or artifact is missing (session.ErrStateKeyNotExist, fs.ErrNotExist). Any other error from the session or artifact service now fails the run. A custom artifact.Service must wrap fs.ErrNotExist for missing files, and an agenttool sub-agent with an optional artifact placeholder now errors when the root runner has no artifact service. (#1748)
• Go 1.26.10 is required: go.mod moves from 1.26.6 to 1.26.10. (#1916)
• Agent Analytics plugin defaults: the default table name is agent_events instead of events, existing tables are upgraded additively on startup unless DisableSchemaUpgrade is set, the per-event views are created on startup unless DisableCreateViews is set, and granular event types (STATE_DELTA, AGENT_TRANSFER, HITL_*, NODE_OUTPUT, …) replace the generic EVENT row. These apply when the plugin module is next tagged, not to the core v2.7.0 tag. (#1786, #1788, #1791)

Smaller behavior changes:

• geminitool.GoogleSearch returns an error on a non-Gemini model, and on a Gemini 1.x model combined with other tools. (#2026)
• launcher: passing telemetry.WithOtelToCloud(true) in TelemetryOptions now exports to Google Cloud and needs credentials at startup, where it used to start with export silently off. (#1968)
• openaimodel: a config field added by a genai release newer than v1.74.0 is refused with ErrUnsupportedConfigField. Configs valid today are unaffected. (#1922)
• mcptoolset with Config.Auth opens one session per distinct credential, closes sessions idle for 10 minutes, and caps the pool at 64. Without Auth nothing changes. (#1755)
• LoopAgent yields the context error once and stops when its context is cancelled, where it used to keep looping if no pass yielded an event. (#1810)
• An agent callback that saves an artifact and returns nil now yields one event carrying the ArtifactDelta, where it yielded none. (#1878)
• The in-memory memory service no longer stores or matches thought parts, and an event with only thought text is not stored. (#1975)
• Deprecated: artifact.SaveRequest.Version. Save has never read it. (#1985)

Full Changelog: v2.6.0...v2.7.0

New Contributors

• @haiyuan-eng-google made their first contribution in #1786
• @venkat-uk made their first contribution in #1412
• @tonydzi made their first contribution in #1543
• @jaideeppyne made their first contribution in #1604
• @beiryu made their first contribution in #1872
• @DisturbedCrow made their first contribution in #1964
• @Saurav-Gupta-9741 made their first contribution in #1968
• @ScottMansfield made their first contribution in #1866
• @rafal-glowinski made their first contribution in #2003