[Release] google/adk-python v2.6.0: v2.6.0

Google ADK Python v2.6.0 introduces significant enhancements for agent development, including per-invocation auth headers for A2A, new evaluation samples, OAuth2 improvements, and telemetry capabilities. Key additions include the ReflectAndRetryModelPlugin for self-healing, OCI Generative AI provider integration, Eventarc Advanced toolset, and ManagedAgent instruction field support. The release also improves CLI functionality with telemetry consent checks, command logging, and enhanced error handling across multiple agent frameworks.

Key Points

• •Per-invocation auth headers support for A2A agent card fetching enables more flexible authentication scenarios
• •New --extra_packages option for `adk deploy agent_engine` allows custom dependency management during deployment
• •ReflectAndRetryModelPlugin provides self-healing capabilities for model errors, improving agent robustness
• •OCI Generative AI provider integration expands LLM options beyond Google's offerings
• •Eventarc Advanced toolset enables event-driven agent architectures with improved event handling
• •ManagedAgent instruction field allows system-level instruction forwarding for better agent control
• •Enhanced telemetry with consent checks, status commands, and CLI execution metrics for better observability
• •OAuth2 improvements including nonce forwarding and token refresh in OpenAPI credential exchanger
• •Multiple ADK evaluation samples (home-automation, basic_criteria, custom_metric, llm_judge_match, rubric_criteria, user_simulation) provide reference implementations
• •Audio generation user simulator and live mode state_delta support expand evaluation capabilities

Workflow Diagram

Concepts

Artifacts (5)

adk deploy agent_engine --extra_packages <package_list>

# ManagedAgent with instruction field
# The instruction field is forwarded as system_instruction to the agent
agent = ManagedAgent(
    name="custom_agent",
    instruction="You are a helpful assistant that..."
)

from adk.agents import ReflectAndRetryModelPlugin

# Add self-healing capability to agents
plugin = ReflectAndRetryModelPlugin(
    max_retries=3,
    error_recovery=True
)
agent.add_plugin(plugin)

# OCI Generative AI Provider Configuration
provider_config = {
    "type": "oci_generative_ai",
    "endpoint": "<oci_endpoint>",
    "credentials": "<oci_credentials>",
    "model": "<model_name>"
}

from adk.telemetry import TelemetryConsent

# Check and manage telemetry consent
consent = TelemetryConsent()
if consent.is_enabled():
    # Telemetry collection is active
    pass

# Configure telemetry endpoints
consent.configure_endpoints()
consent.write_local_config()