Google Cloud Solution Agentic AI Bidirectional Streaming

Live bidirectional multimodal streaming agentic AI solution

This skill guides agents through the workflow to design and implement a tailored multi-product solution in the cloud for a live, bidirectional multimodal streaming workload, use case, or requirement.

Workflow

The solution design and implementation workflow consists of the following phases:

• Phase 1: Requirements discovery and analysis: Analyze the workload's requirements, constraints, dependencies, and current state.
• Phase 2: Solution design: Build a technology stack, architecture, and deployment configuration for the workload based on Google Cloud design best practices and recommendations.
• Phase 3: Implementation plan: Generate automation and instructions to deploy the solution.
• Phase 4: Solution validation: Validate that the deployment meets the requirements of the workload.

Supporting links

Use these references to answer user questions that might deviate from the workflow phases.

• Architecture guides:
  • Enable live bidirectional multimodal streaming
  • Choose your agentic AI architecture components
  • Multi-agent AI system in Google Cloud
  • Choose a design pattern for your agentic AI system
  • Multi-agent private networking patterns in Google Cloud
• Implementation guides:
  • Host AI agents on Cloud Run
  • Triggering Cloud Run with WebSockets
  • Start and Manage a Gemini Live API Session
  • ADK Streaming Tools
  • ADK Streaming Configuration
  • Codelab: Way Back Home Level 4 instructions (and solution code)

Phase 1: Requirements discovery and analysis

1. Discover requirements: Understand the functional and non-functional requirements, business goals, and current state (if any) of the workload, including its architecture, dependencies, and constraints. Use the following questions to guide the requirements discovery process:

   • What are the primary input modalities (audio, video, or text) and what is the target latency for real-time, narrated feedback?
   • Do you require real-time safety monitoring, hazard detection, or visual inspection? If so, then what specific safety hazards, operational risks, or incorrect steps need to be monitored and detected in the video stream?
   • What existing systems, knowledge bases, product documentation, or schematic repositories must the AI agents access for grounded guidance?
   • What are the client-side device constraints and network limitations?

2. Identify components: Based on the requirements analysis, identify the components of the workload and their relationships. Also identify any cross-cloud components, hybrid components, or on-prem components that the solution needs to integrate with.

3. Generate component decomposition: Generate a technical decomposition of the components of the workload.

4. Ask for confirmation: Ask the user to confirm whether the generated technical decomposition matches their workload requirements.

5. Iterate: If the user requests changes, then generate an updated technical decomposition, and ask the user to confirm the changes. Continue iterating until the user confirms the technical decomposition.

Phase 2: Solution design

1. Retrieve relevant Google Cloud documentation:

   • Enable live bidirectional multimodal streaming
   • Multi-agent AI system in Google Cloud
   • Choose your agentic AI architecture components
   • Multi-agent private networking patterns in Google Cloud

   Important: Use the content that you retrieve from Google Cloud documentation to ground the guidance that you generate in the remaining steps of this phase.

2. Map components to Google Cloud products:

   • Cloud Run networking:
     • Recommended primary configuration: Regional External Application Load Balancer combined with Cloud Armor for HTTP/HTTPS/WebSocket ingress, and Direct VPC egress for Cloud Run private network access.
     • Alternative product 1: Global External Application Load Balancer
       • Pros: Single anycast IP, global IPv6 termination, and low-latency routes to globally distributed backend services.
       • Cons: Terminates TLS globally at edge locations, which might not comply with strict regional data residency regulations.
     • Alternative product 2: Internal Application Load Balancer
       • Pros: Securely exposes Cloud Run services internally within the VPC to meet internal ingress criteria, terminates TLS with trusted certificates, and supports Cloud Armor backend security policies.
       • Cons: Requires that you configure serverless network endpoint groups (NEGs) as backends and manage load balancer resources.
     • Alternative product 3: Private Service Connect interface
       • Pros: Secure private VPC connections for Gemini Enterprise Agent Runtime that uses network attachments.
       • Cons: Limited to RFC 1918 routable subnet ranges, requires proxy setup for non-routable/internet destinations.
   • Frontend:
     • Recommended primary product: Cloud Run
     • Alternative product 1: Firebase App Hosting
       • Pros: Automated builds and deployment pipeline from GitHub, optimized for modern framework integrations.
       • Cons: Less control over container configurations, limits customization of low-level networking.
     • Alternative product 2: Google Kubernetes Engine (GKE)
       • Pros: Maximum control over routing, scaling, and custom container runtimes.
       • Cons: Significant infrastructure management complexity and cost overhead.
   • Agent development framework:
     • Recommended primary product: Agent Development Kit (ADK).
   • Agent-to-agent communication:
     • Recommended primary product: Agent2Agent (A2A) protocol.
   • Runtime for your agent:
     • Recommended primary product: Cloud Run
     • Alternative product 1: Gemini Enterprise Agent Runtime
       • Pros: Fully managed Python runtime, built-in memory/sessions, secure code execution sandbox.
       • Cons: Limited to Python, does not support hosting custom MCP servers, less control over container environment.
     • Alternative product 2: Google Kubernetes Engine (GKE)
       • Pros: Maximum infrastructure control, stateful pods, custom scaling.
       • Cons: High operational complexity and overhead.
   • Model runtime:
     • Recommended primary product: Gemini Enterprise Agent Platform
     • Alternative product 1: Cloud Run
       • Pros: Serverless hosting for containerized open/custom models like Gemma.
       • Cons: Cannot serve Google Gemini models, manual instance scaling overhead.
     • Alternative product 2: Google Kubernetes Engine (GKE)
       • Pros: Maximum control over inference server on GPU/TPU nodes, cheap for predictable high volume.
       • Cons: Cannot run Google Gemini models, high cluster management overhead.
   • Model selection:
     • Recommended primary product: Gemini Flash with Gemini Live API
     • Alternative product 1: Gemini Pro
       • Pros: Highest capability for reasoning, complex instructions, context tracking, and multi-agent coordination.
       • Cons: Higher request cost and latency, which makes it less suitable for real-time conversational requirements.
   • VPC connection to the database: The architect agent sends queries through this connector to securely access resources in the Virtual Private Cloud (VPC) network used for storage resources in this architecture.
     • Recommended primary product: Serverless VPC Access connector.
     • Alternative product 1: Direct VPC egress
       • Pros: Lower latency, lower resource cost, and avoids throughput scaling bottlenecks.
       • Cons: Requires specific routing and subnet configurations.
   • Caching:
     • Recommended primary product: Memorystore for Redis Cluster
   • Database:
     • Recommended primary product: Google Cloud Databases Use the recommendations listed on this page to help the user choose the appropriate database option.
     • Alternative product 1: Compute Engine (for self-hosted databases)
       • Pros: Full control over database configurations, OS accessibility, and custom database engines or extensions.
       • Cons: High operational overhead to manually manage backups, patching, scaling, and high availability.

3. Create architecture diagram: Create an architecture diagram that shows the components, their relationships, and data and control flows.

   • The diagram must be in the Mermaid format: https://github.com/mermaid-js/mermaid.
   • The diagram must use Google-approved icons based on the guidance in https://services.google.com/fh/files/misc/google-cloud-product-icons.pdf.

4. Generate design recommendations: Generate design guidance based on the following Google Cloud best practices and recommendations:

   • Security, privacy, and compliance:
     • To limit access to the app, disable the default run.app URL of the frontend Cloud Run service and configure a regional external Application Load Balancer with Cloud Armor security policies to handle request filtering, rate limiting, and DDoS protection.
     • Enforce the principle of least privilege when you configure IAM permissions for resources in the topology.
     • To protect sensitive multimodal data (such as voice prints and video), enforce TLS encryption for all bidirectional WebSocket connections.
     • Secure Agent2Agent (A2A) communication using authenticated extended agent cards, and attach OpenID Connect (OIDC) identity tokens to let IAM validate that only authorized agents access the data.
     • Incorporate human-in-the-loop flows to let supervisors monitor, pause, and override business-critical agent actions.
     • For more information about security considerations, see https://docs.cloud.google.com/architecture/framework/perspectives/ai-ml/security.md.txt
   • Reliability:
     • Build fault-tolerant agents employing decentralized designs where agents can operate independently to survive failures.
     • Simulate inter-agent coordination issues and unexpected behaviors in a replica staging environment before deploying to production.
     • Leverage regional multi-zone deployment of Cloud Run to automatically load-balance and survive zone outages.
     • Plan model capacity by monitoring standard quota rates, using Provisioned Throughput for business-critical production workloads.
     • For more information about reliability considerations, see https://docs.cloud.google.com/architecture/framework/perspectives/ai-ml/reliability.md.txt
   • Operational excellence:
     • Route agent logs to Cloud Logging in structured formats, integrating standard stdout/stderr streams.
     • Track complete agent workflows, reasoning loops, and execution paths using Cloud Trace and trace visualizers.
     • Perform continuous evaluation using tools like Agent Evaluation on Gemini Enterprise Agent Platform or ADK evaluation methodologies.
     • Centralize database tools and connection scaling policies using the MCP Database Toolbox.
     • For more information about operational excellence considerations, see https://docs.cloud.google.com/architecture/framework/perspectives/ai-ml/operational-excellence.md.txt
   • Cost optimization:
     • Reduce data ingestion costs by employing low-frequency frame sampling and compressing video to Base64 JPEGs.
     • Use context caching for requests containing long system prompts or static lookup databases to reduce input token costs.
     • Structure prompts to get concise responses to minimize generation token outputs.
     • Start with the most smaller and cost-efficient models. Upgrade to more powerful models with reasoning based on performance requirements.
     • For more information about cost optimization considerations, see https://docs.cloud.google.com/architecture/framework/perspectives/ai-ml/cost-optimization.md.txt
   • Performance efficiency:
     • Decouple incoming audio and video packets from the model's inference engine. Use a thread-safe, asynchronous First-In-First-Out (FIFO) buffer to keep the user interface responsive to interruptions.
     • To achieve sub-millisecond read speeds and prevent silences during real-time voice interactions, deploy an in-memory Memorystore for Redis Cluster database for the agent's schematic vault.
     • To optimize service performance, configure memory limits and CPU limits allocated to the Cloud Run instances based on live workloads.
     • For more information about performance considerations, see https://docs.cloud.google.com/architecture/framework/perspectives/ai-ml/performance-optimization.md.txt.
   • Sustainability:
     • Route simpler tasks to small language models (SLMs) and optimize model routing to minimize total model inference footprint.
     • To prevent wasting resource baseline energy, use Cloud Run native autoscaling to scale compute runtimes down to zero during idle periods.
     • For more information about sustainability considerations, see https://docs.cloud.google.com/architecture/framework/sustainability/printable.md.txt.

5. Draft solution architecture: Compile the requirements, technical decomposition, product mapping, architecture diagram, and design recommendations into a single Markdown file.

6. Request review: Present the generated solution architecture to the user and request their feedback or approval.

7. Iterate: If the user requests changes, generate an updated solution architecture and repeat steps 5-6 until the user approves the solution architecture.

Phase 3: Implementation plan

1. Retrieve relevant implementation resources:

   • Host AI agents on Cloud Run
   • Triggering Cloud Run with WebSockets
   • Start and Manage a Gemini Live API Session
   • ADK Streaming Tools
   • ADK Streaming Configuration
   • Codelab: Way Back Home Level 4 instructions (and solution code)

   Important: Use these resources as the technical foundation for the IaC and deployment instructions you generate in the remaining steps of this phase.

2. Identify deployment prerequisites: Document prerequisites for the deployment, including the following:

   • Projects and billing associations
   • Required Google Cloud APIs
   • Required IAM permissions
   • Any other prerequisites

3. Generate Infrastructure as Code (IaC): Generate code, like Terraform, and deployment scripts to automate the provisioning of the proposed Google Cloud resources.

4. Write deployment instructions: Draft sequential, step-by-step deployment instructions to execute the IaC and initialize the workload components.

5. Request review: Present the generated deployment instructions to the user for feedback and confirmation.

6. Iterate: If the user requests changes, then generate an updated implementation plan and repeat steps 4-5 until the user approves the implementation plan.

Phase 4: Solution validation

1. Retrieve relevant verification resources:

   • Host AI agents on Cloud Run
   • Triggering Cloud Run with WebSockets
   • Start and Manage a Gemini Live API Session
   • ADK Streaming Tools
   • ADK Streaming Configuration
   • Codelab: Way Back Home Level 4 instructions (and solution code)

   Important: Use these resources and their verification patterns as the starting point for the validation checks and verification scripts that you generate in the remaining steps of this phase.

2. Define validation checks: Outline validation steps to verify that the deployed infrastructure meets the workload's requirements:

   • Deployment dry-run: Commands like terraform plan to preview changes.
   • Connectivity and routing: Verification of network paths, load balancer routing, and service endpoints.
   • Security policies: Verification of restricted access, firewall rules, and IAM enforcement.

3. Generate verification scripts: Draft lightweight scripts or command-line instructions, such as using curl or gcloud, that the user can run to perform these validation checks.

4. Compile validation report: Document the validation steps, verification scripts, and expected outcomes in a single Markdown file.

5. Conduct validation and finalize: Assist the user in executing the validation checks and troubleshooting any deployment issues. After the solution is validated successfully, request final approval from the user.

6. Iterate: If the user requests changes, then generate an updated validation plan and repeat steps 4-5 until the user approves the validation plan.