Conversational Analytics API Implements A2A Protocol for Multi-Agent Data Orchestration

Google has launched the Conversational Analytics API to enable multi-agent systems to discover agent capabilities, delegate analysis queries, and stream structured responses like SQL queries and Vega-Lite chart specifications, docs.cloud.google.com reported. The new application programming interface implements the open agent-to-agent protocol for Google Cloud. Developers can query built-in data agents or set up custom agents to handle enterprise business logic.

Google Cloud Deploys Conversational Analytics API for Multi-Agent Workflows

The Conversational Analytics API provides foundational plumbing for modern multi-agent systems. Orchestrator agents inspect target agent cards before delegating tasks to understand available skills and descriptions. Applications can target built-in data agents or route traffic to custom configurations based on domain requirements.

To query built-in data sources, the system relies on specific resource paths. Developers specify agents/bigquery-ca for BigQuery datasets or agents/looker-ca for Looker Explore environments. Custom data agents use individual identifiers structured as dataAgents/DATA_AGENT_ID.

Engineering Teams Configure Google Cloud Environment and IAM Roles

Before implementing the API, engineering teams must configure access within their Google Cloud environment. Administrators need to enable the Conversational Analytics API, BigQuery API, and Looker API. Client libraries must be installed and authentication tokens acquired for the API endpoints.

IAM roles govern permission boundaries across the workflow. Administrators must assign the Gemini Data Analytics Data Agent User role, designated as roles/geminidataanalytics.dataAgentUser, or the stateless equivalent roles/geminidataanalytics.dataAgentStatelessUser. Targets require standard data access, such as the roles/bigquery.dataViewer permission for underlying BigQuery data sets.

Orchestrators Retrieve Agent Metadata Using the getCard Method

Orchestrators inspect agent capabilities using the getCard method prior to query delegation. This method retrieves metadata including descriptions, supported extensions, and available skills. The protocol applies uniformly across built-in and custom configurations.

Python developers access these features by importing the software development kit from google.cloud import geminidataanalytics_v1. Code implementations substitute the PROJECT_ID with the Google Cloud project identifier and specify the LOCATION parameter, such as us, us-east4, eu, or global, depending on resource deployment.

Query Execution Yields Complete Responses or Streaming Updates

Query execution yields either complete responses or streaming updates. Streaming outputs deliver inference progress alongside structured artifacts, including executable SQL statements and Vega-Lite chart specs. Python and HTTP software development kits support direct agent construction and querying.