Google ADK 2.8.0 adds a Model Armor guardrail plugin and fences relayed sub-agent output so it cannot pose as instructions

GitHub (google/adk-python)·medium signalXBlueskyLinkedInCopy link

ADK for Python 2.8.0 landed 26 August with a Model Armor guardrail plugin, a fix that fences relayed agent output so a sub-agent's text cannot be read as instructions by its caller, SQL injection guards in the BigQuery tools, blocked yaml and ruamel deserialization in agent-config code references, and auth taken from the request rather than the client's response. It also adds native task mode for RemoteA2aAgent with auth_scheme and auth_credential support, an ADK_MAX_LLM_CALLS environment variable, and an opt-in experimental telemetry layer reporting per-invocation and per-workflow token spend and tool call counts behind ADK_EXPERIMENTAL_TELEMETRY=true.