Skip to content
Agents tracked: 258 Downloads (7d): 219M up 6.1% GitHub stars: 5.5M VS Code installs: 148M Releases (7d): 293 Agent status: 2 with issues Updated Oct 7, 2026
AgentGid

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. 3 months covered; the current month updates daily.

October 2026 · so far

1 release: 2.11.0

October brought significant enhancements to execution control, workflow approval handling, and agent capabilities. The release adds graceful cancellation, tool confirmation workflows, and new options for agent consultation and memory storage.

Highlights

  • Execution cancellation now available through AbortSignal for Runners, Workflows, and nodes, with automatic cancellation when SSE clients disconnect.
  • Tool nodes in workflows can now pause for user approval via RequestInput instead of passing errors downstream.
  • New ModelConsultTool lets agents consult another model mid-task with configurable per-turn and per-session budgets.
  • SQLite memory service option for storing agent memory in a local database.
  • Dev UI runtime configuration now served dynamically by the server instead of from a static file.

New

  • Execution cancellation via AbortSignal for graceful run termination
  • Tool confirmation workflow using RequestInput for user approval
  • ModelConsultTool for mid-task consultation of other models with budget limits
  • SQLite memory service backend with sqlite:// URI support
  • MCP SDK 2.x support for connecting to modern protocol MCP servers
  • Sample memory profile preloading instructions

Improved

  • Dev UI runtime configuration served dynamically per request instead of written to package
  • SSE run endpoint (/runsse) now properly cancels runs when client disconnects

September 2026

4 releases: 2.9.0 → 2.10.0

September brought significant enhancements to model reliability, integrations, and workflow capabilities. The month focused on expanding agent development options and improving execution safety.

Action needed
  • GCS tool local file access requires localfileroot setting; access is refused entirely when absent.
  • Workflow node behavior changed: failed nodes now re-run on resume instead of replaying as completed; ensure node bodies are idempotent.
  • InMemorySessionService now raises SessionNotFoundError for append operations on unknown sessions instead of accepting and discarding events.

Highlights

  • New LiveKit runner enables building rich voice and telephony agents.
  • Automatic model failover protects agent uptime when primary models encounter errors.
  • YAML-based workflow configuration simplifies authoring and loading of graph workflows.
  • MongoDB toolset adds high-performance vector and hybrid search capabilities.
  • New evaluation metrics help optimize agent execution costs and latency.

New

  • LiveKit runner for voice and telephony agent development.
  • Model failover system that automatically switches to backup models on errors.
  • YAML configuration support for ADK 2.0 graph workflows.
  • MongoDB integration with vector and hybrid search tools.
  • Evaluation metrics for tracking duration, token consumption, and model call counts.
  • Ephemeral lifecycles and active skill limits for dynamic resource control (experimental).
  • OpenAI reasoning model parameter adaptation and precise reasoning token reporting.
  • Support for MCP SDK 2.x servers alongside existing 1.x compatibility.

Improved

  • Workflow node resumption now re-runs failed nodes instead of replaying them as completed.
  • GCS tools now restrict local file access to a specified directory only.
  • Claude adaptive thinking tasks now expose visible step-by-step reasoning.
  • OpenTelemetry event names are preserved on platforms outside Agent Engine.
  • BigQuery protected write mode validates statements against session dataset.
  • BigQuery protected mode stores sessions in process memory to isolate temporary tables.
  • Nested workflow retries now work when a node inside fails.

Fixed

  • InMemorySessionService now raises SessionNotFoundError instead of silently discarding events for unknown sessions.
  • OpenAIResponsesLlm no longer ignores thinking configuration silently; now logs a warning and directs users to OpenAIGenerateContentConfig.effort.
  • AgentEvaluator.evaluate and evaluateevalset now raise ValueError when no evaluation cases are processed, 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.

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