Google ADK Python v2.8.0

Google ADK Python v2.8.0 delivers several major features and enhancements across agent orchestration, tooling, live sessions, and telemetry. Key updates include: - Agent-to-Agent (A2A) : Added native task mode support, auth scheme/credential configuration, and auth resolution in AgentRegistry for RemoteA2aAgent . Added sub-agent escalation events to ParallelAgent . - Security & Guardrails : Fixed SQL injection vulnerabilities in BigQuery tools. Added the Model Armor guardrail plugin. - Live Streaming & Modalities : Added VIDEO to allowed modalities in the ADK API server. Added direct messaging to users from live streaming tools, support for pre/post-model callbacks in live mode, live interaction status on LlmResponse and Event , and session resumption via RunConfig.session resumption handle . - LLM Integrations & Tooling : Supported injecting custom LLM clients into Gemini and AnthropicLlm . Added ADK MAX LLM CALLS environment variable to cap LLM calls. Allowed streaming tools to schedule response handling per yield. Added data agent management tools (create, delete, update, list with location) and spreadsheet MIME support to load artifact . - Evaluation & Telemetry : Added custom metrics and parallelized LLM-as-judge evaluation in AgentEvaluator . Added telemetry for per-workflow and per-invocation token spend, workflow tool/inference counts, MCP HTTP logging via OpenTelemetry, context cache states, and load skill spans. - Platform & Ecosystem : Added Cloud Build worker pool support for Agent Engine deployments, configurable Vertex AI API versions, memory ID/allowed…

• Update to Google ADK Python v2.8.0 to resolve SQL injection vulnerabilities in BigQuery tools.
• Optionally configure the new ADK MAX LLM CALLS environment variable to guard against unbounded LLM call loops.
• Review custom LLM client integrations for Gemini and AnthropicLlm, or adopt RemoteA2aAgent task mode and auth improvements if using agent-to-agent architectures.