Architecture for Scalable Agentic Workflows

The system replaces manual project vetting with an automated pipeline that enriches incoming requests with risk scores, resource availability, and historical analysis. The architecture is built on three distinct layers:

  • Interface Layer: Asana serves as the human-facing input. A form submission triggers a webhook that initiates the entire processing chain.
  • Orchestration Layer: A central orchestrator built with the Google Agent Development Kit (ADK) manages the workflow. ADK simplifies agent development by handling session state, tool dispatching, and Gemini API integration, allowing the developer to define agent behavior through simple class-based instructions.
  • Specialist Agent Layer: Four independent agents (Risk Scorer, Resource Advisor, Asana Context, and BigQuery Analyst) execute tasks. These agents run as individual Cloud Run containers, which scale to zero when idle, ensuring cost-efficiency by charging only per request.

Leveraging Open Standards and Parallel Execution

The system utilizes the Agent-to-Agent (A2A) protocol to decouple the orchestrator from the specialist agents. The orchestrator does not need to understand the internal logic of each specialist; it simply dispatches tasks to specific URLs and waits for results. This modularity allows for swapping or updating agents without modifying the orchestrator code.

To avoid the limitations of static data injection (which risks context window overflow and stale data), the system uses the Model Context Protocol (MCP) to connect agents directly to live data sources:

  • BigQuery Integration: The BigQuery agent dynamically generates SQL at runtime to query historical project data, enabling pattern recognition across years of enterprise records.
  • Live Work Graph: The Asana MCP server allows agents to query live workspace context, such as current team task loads and overlapping project timelines, ensuring decisions are based on real-time data rather than snapshots.

Observability and Security

By deploying agents as containerized services on Google Cloud, the system benefits from native integration with Google Cloud Logging. Every step—including agent payloads, tool calls, and inter-agent communication—is traceable. Security is maintained through the use of Application Default Credentials and Secret Manager, ensuring that agents interact with enterprise data sources securely without hardcoding sensitive information.