Build Production-Grade AI Agents with ADK

Mandatory primer bundled with every track. Sourced from Google Skills #1674 (AI Boost Bootcamp) + #1401. Covers what Gemini Enterprise is (vs Gemini for Workspace vs the Gemini Enterprise Agent Platform, the platform formerly known as Vertex AI); GCP basics; the agentic lifecycle (Build → Scale → Govern → Optimize); and terminology alignment (agents, RAG, grounding, ADK, GEAR). Completion check before Day 1.

Agentic AI vs traditional AI. Agent anatomy: model, tools, memory, system instructions. Reasoning loops (gather → act → verify). Where agents fit in Google Cloud (Gemini Enterprise Agent Platform, Gemini API, ADK, Agent Engine). Mantra: Develop with ADK; Operate with Gemini Enterprise.

ADK languages (Python, TypeScript, Go, Java). Core abstractions: Agents, Tools, Workflow Agents, Memory Bank. Local dev setup; project scaffolding. Tool ecosystem (3rd-party + custom tools). Predictable pipelines vs agent-coordinated routing. Built-in evaluation framework. Deployment surfaces (Local, Agent Runtime, Cloud Run, GKE).

Parent-child agent flows; task delegation patterns. Model Context Protocol (MCP) for secure enterprise system connections. Agent2Agent (A2A) Protocol for multi-agent communication. Workflow Agents controlling work-flow automatically. Session state propagation between agents. Compliance-focused deterministic flows for critical ops.

Visual workflow builder for non-developers. Pre-built templates via Agent Garden (code modernisation, financial analysis, invoice processing). Direct export from Agent Studio to ADK for deeper customisation. When to choose Agent Studio vs Agent Designer vs ADK.

Hands-on lab using the Cymbal Foods Marketing scenario (Skills #1586). Prepare data, create the GE application, connect data stores (GCS, Drive, Calendar), interact with agents, use the Deep Research agent, focused NotebookLM analysis. Earn the 'Deploy an Agent with ADK' skill badge.

RAG (Retrieval-Augmented Generation) fundamentals and tooling. RAG in action lab. Search agents: power and limits. Gemini Enterprise Agent Platform Search shopping-experience use case. Build-your-own-agent walk-through. Bridges this track to the CX track.

Gemini Enterprise Agent Platform Studio (formerly Vertex AI Studio) walkthrough. Gemini API,2.0 Flash, multimodal reasoning. Model selection by complexity (2.5-flash vs 2.5-pro) for cost / performance trade-off. Hands-on labs designed for student / developer curriculum use. Token budgeting per project; cost tracking at scale.

Deployment targets: Gemini Enterprise Agent Engine (managed), Cloud Run (containerised), GKE (full control). Containerising tool-using agents. Agent Runtime with sub-second cold starts. Long-running agent support (multi-day workflows). Bidirectional streaming (WebSocket). Live audio and video via multimodal streaming. Skill badge: Deploy an Agent with ADK challenge.

Agent Identity gives a cryptographic ID per agent plus audit trails. Agent Registry is the single source of truth. Agent Gateway is the 'air traffic control' for agent-to-tool interactions. Agent Anomaly + Threat Detection via Security Command Center integration. CISO-approved guardrails. Agent Sandbox provides hardened execution for model-generated code.

Agent Simulation provides synthetic users + virtualised tools for pre-prod testing. Agent Evaluation runs multi-turn autoraters with turnkey dashboards. Agent Observability captures execution traces and real-time debugging. Agent Optimizer performs automatic failure clustering and refines system-instruction suggestions. KPIs and alert configuration.

Operational rollout pattern for 20K to 50K Gemini Enterprise Agent Platform projects. GUI front-end for admin staff (Terraform-backed). Auto-provisioning + offboarding; bulk operations. Budget alerts and auto-terminate; student access revocation. Memory Bank for per-user personalisation. Agent Sandbox for safe student-code execution. Cost attribution per department / faculty / course. Decommissioning at semester end.

Individual portfolio piece. Build, deploy, and evaluate a production-ready agent for a real business use case. Choose from instructor-supplied scenarios: FinTech compliance agent · IT services automation · BPO workflow agent · banking document intelligence · autonomous-enterprise pattern, or bring your own. Wire connectors, agents, eval, observability, and a deployment surface (Agent Engine or Cloud Run). Peer review on Day 5. Output goes into your GitHub / portfolio.

Reference architecture for deploying agents into a Gemini Enterprise org: where your agent sits relative to the connector mesh, identity plane, Agent Gateway, and consumption surfaces. Phased rollout patterns (pilot → department → enterprise), environment promotion (dev → stage → prod), DR + multi-region posture for Agent Engine vs Cloud Run, and the decision tree for landing-zone choices. Walk away with a deployment diagram you can hand to a platform team.

Cloud Logging for agent workloads. Agent step logs, tool-call logs, model-call logs, and Audit Logs, what each captures, retention defaults, and structured-logging conventions for agent telemetry. Log sinks to BigQuery + GCS for replay and offline eval. Log-based metrics for agent KPIs (success rate, tool-error rate, latency percentiles). PII-aware log scrubbing. The exact log queries you'll run during the first prod incident.

Cloud Trace + OpenTelemetry + ADK trace viewer for end-to-end agent tracing across reasoning, tool calls, retrieval, and model invocations. Reproducing a failed agent run from trace ID to root cause. Common failure modes: tool schema drift, hallucinated arguments, context-window overflow, retrieval miss, infinite tool loop, deadlock between agents. Live debugging walkthrough on real agent traces. The reproduce → bisect → fix loop.

FinOps for agent workloads. Token economics per agent step; per-conversation cost attribution. Model routing (Flash for cheap steps, Pro for reasoning, Ultra only when needed). Context caching and prompt-size discipline as the two biggest token-cost levers. Tool-call batching and parallel-fanout patterns. Budget alerts + auto-throttle. The cost-per-successful-task metric that matters more than raw token spend.

End-to-end LLM Ops lifecycle on Gemini Enterprise. Prompt versioning and A/B rollout via Prompt Management. Eval pipelines with Agent Evaluation (autoraters, golden datasets, regression gates). Shadow traffic + canary releases for new model versions or system-prompt changes. Drift detection on inputs, outputs, and tool-call patterns. Continuous improvement loop wired through Agent Optimizer. The agent-equivalent of CI/CD.

The Day-2 playbook for agent owners. Top 20 production failures we see across engagements, each with symptom, root cause, and the exact fix: agent stuck in tool loops, schema-validation rejections, grounding miss, context overflow, ACL leakage into outputs, latency-tail spikes, eval-prod divergence, quota exhaustion, silent regressions after a model upgrade. Reference card you'll keep open during the first 90 days post-launch.