PRD: ADK Deployment Specialist
Version: 2.1.0 Author: Jeremy Longshore jeremy@intentsolutions.io Status: Active Marketplace: tonsofskills.com by Intent Solutions Portfolio: jeremylongshore.com
Problem Statement
Deploying ADK multi-agent systems to Vertex AI Agent Engine involves coordinating multiple complex surfaces: agent orchestration patterns (Sequential/Parallel/Loop), A2A protocol endpoints, Code Execution Sandbox configuration, Memory Bank state management, and IAM/networking setup. Getting any one of these wrong causes silent failures, broken inter-agent communication, or security vulnerabilities. Developers spend hours debugging deployment issues that stem from misconfigured agent contracts or missing A2A endpoints.
Target Users
| User | Context | Primary Need |
|---|---|---|
| AI Engineer | Building a new multi-agent system for production deployment | End-to-end deployment from local agent code to live Agent Engine endpoints |
| Platform Engineer | Migrating existing agents from local dev to Agent Engine | Reliable deployment pipeline with endpoint validation and rollback guidance |
| DevOps Engineer | Setting up CI/CD for agent deployments | Automated deployment commands with health checks and observability hooks |
Success Criteria
- Deploy an ADK agent to Agent Engine with verified A2A endpoints in under 15 minutes
- All A2A protocol endpoints (AgentCard, task send, task status) respond correctly post-deployment
- Code Execution Sandbox configured with 7-14 day TTL and SECURE_ISOLATED sandbox type
- Memory Bank enabled with minimum 100-memory retention and Firestore encryption
- Deployment includes observability: structured logging, retry/backoff, and health monitoring
- Agent package excludes test files and dev dependencies (minimized cold start)
Functional Requirements
- Confirm the desired architecture (single vs multi-agent) and orchestration pattern (Sequential/Parallel/Loop)
- Define AgentCard and A2A interfaces: inputs, outputs, task submission, and status polling contracts
- Implement agent(s) with the minimum required tool surface including Code Execution Sandbox and Memory Bank as needed
- Test locally with representative prompts and failure cases, then generate smoke tests for post-deploy
- Deploy to Vertex AI Agent Engine using the Python SDK (
vertexai.Client.agent_engines.create()) - Validate deployed endpoints:
/.well-known/agent-card,POST /v1/tasks:send,GET /v1/tasks/<id> - Configure observability: structured logs, Cloud Monitoring dashboards, and retry/backoff for transient failures
Non-Functional Requirements
- All deployments use OIDC/WIF for authentication; never commit long-lived service account keys
- Agent packages must exclude test files and dev dependencies to minimize cold start time
- Deployment commands must be idempotent (safe to re-run without side effects)
- Support for both greenfield deployments and updates to existing Agent Engine instances
- Local tests must pass before any deployment attempt (fail-fast principle)
- All generated code must include error handling for transient failures (retries with backoff)
- Deployment scripts must provide clear rollback instructions if validation fails
Dependencies
- Google Cloud project with Vertex AI API enabled and Agent Engine permissions
- ADK installed and pinned to the project's supported version
- Python SDK
google-cloud-aiplatform[agent_engines]>=1.120.0 gcloudCLI authenticated with deployment permissions- A test runner (pytest) available in the repository
Out of Scope
- Infrastructure provisioning with Terraform (handled by adk-infra-expert)
- Post-deployment inspection and scoring (handled by vertex-engine-inspector)
- CI/CD pipeline creation for GitHub Actions (handled by gh-actions-validator)
- Cost optimization and model selection strategy
- Agent application logic design (handled by adk-engineer)
- Multi-region deployment with traffic splitting