Automate project intake with multi-agent AI using MCP, Google ADK, Cloud Run, and BigQuery

Key Concepts

• Multi-Agent System (MAS): A distributed architecture where specialized AI agents perform specific tasks and communicate to solve complex problems.
• Google ADK (Agent Development Kit): A framework that simplifies agent orchestration, session state management, and tool dispatching.
• MCP (Model Context Protocol): An open standard that allows AI models to connect to external data sources (like BigQuery or Asana) consistently.
• Agent-to-Agent Protocol: A communication pattern where an orchestrator delegates tasks to specialist agents via URLs without needing to understand their internal logic.
• Vertex AI: The platform providing the underlying Gemini models for the agents.
• Cloud Run: A serverless compute platform that scales to zero, ensuring cost-effectiveness by only charging for active request processing.

---

1. System Architecture and Workflow

The system is designed to automate project intake vetting, replacing manual review with an AI-driven pipeline. It operates in three distinct layers:

• Human Interface (Asana): Acts as both the input mechanism (intake form) and the output destination (enriched task).
• Orchestration Layer (Google Cloud): A webhook receiver triggers an ADK orchestrator, which manages the workflow logic.
• Specialist Agent Layer: Four independent agents (BigQuery Analyst, Resource Advisor, Asana Context, and Risk Scorer) execute specific tasks in parallel or sequence.

Step-by-Step Process:

1. Submission: A user submits a project request via an Asana form.
2. Trigger: Asana fires a webhook to the Google Cloud-based orchestrator.
3. Orchestration: The ADK orchestrator receives the task and dispatches specialist agents.
4. Data Retrieval: Agents use MCP servers to query BigQuery (historical data) and Asana (live workspace context) in real-time.
5. Analysis: Agents process the data using Gemini models to generate a risk score, resource check, and recommendation.
6. Enrichment: The final output is pushed back into the original Asana task.

2. Technical Implementation and Frameworks

• ADK Orchestration: The orchestrator uses the ADK to handle session state and tool dispatching. This eliminates the need for manual session management or custom tool-calling logic.
• MCP Integration: By using the BigQuery and Asana MCP servers, the system avoids the "naive approach" of dumping data into a prompt. Instead, agents write their own SQL at runtime, ensuring data is fresh and context windows remain optimized.
• Observability: All interactions—requests, payloads, and agent-to-agent communications—are logged in Google Cloud Logging, providing full traceability for enterprise auditing.
• Security: The system utilizes Application Default Credentials and Secret Manager to ensure secure access to enterprise data.

3. Key Arguments and Perspectives

• Efficiency vs. Manual Vetting: Kevin Blanco argues that manual vetting is prone to error and bottlenecks. By automating the "data work," humans can focus on setting policy while AI handles the heavy lifting.
• Modularity: The use of the agent-to-agent protocol allows for a "plug-and-play" architecture. Specialist agents can be swapped or updated without modifying the orchestrator, as the orchestrator only requires a URL and a standard interface.
• Cost-Effectiveness: By leveraging Cloud Run, the system scales to zero when not in use, ensuring that enterprises only pay for the compute resources consumed during the vetting process.

4. Notable Quotes

• "This isn't a vision of the future. It's a new standard." — Kevin Blanco, regarding the shift toward automated, agent-driven enterprise workflows.
• "Humans set the policy and agents do the data work in seconds with full traceability." — Summarizing the core philosophy of human-AI collaboration.

5. Synthesis and Conclusion

The system demonstrates a shift from monolithic AI applications to modular, agentic workflows. By combining Google ADK for orchestration, MCP for standardized data access, and Cloud Run for serverless execution, the solution solves the problem of fragmented enterprise data. The primary takeaway is that enterprises can achieve high-speed, accurate project vetting by integrating AI directly into existing workflows (Asana) while maintaining security, observability, and cost-efficiency. The architecture is designed to be extensible, allowing organizations to scale their AI capabilities as their project intake needs grow.