3.0 Domain 3 Roadmap: Developing Custom Agents (33% Weight)

Core Architectural Topics in Domain 3 (~33%)

1. Single-Agent ADK Parameters (Module 3.1)

Master description vs. instruction routing, output_schema vs. output_key state handoffs, include_contents, and Gemini Flash vs. Pro model selection.

2. Multi-Agent Orchestration Patterns (Module 3.2)

Design deterministic pipelines with SequentialAgent, ParallelAgent (unique output_key isolation), LoopAgent (escalate=True), and dynamic LlmAgent or AgentTool routing.

3. MCP Toolbox for Databases & Custom Tools (Module 3.3)

Configure tools.yaml (sources, tools, toolsets) for pooled database access, write typed FunctionTools with ToolContext, and use OpenAPIToolset and LongRunningFunctionTool.

4. Agent Search, RAG Engine & Context Caching (Module 3.4)

Choose between turnkey Agent Search (with document ACLs), custom RAG Engine (chunking and semantic reranking), and Gemini Context Caching for large static tokens.

5. State Prefixes, Memory Bank & ArtifactService (Module 3.5)

Scope variables with temp:, session, user:, and app: prefixes, extract cross-session facts in Memory Bank, and version binary files in GcsArtifactService.

6. Agent2Agent (A2A) Interoperability (Module 3.6)

Publish and consume A2A Agent Cards (/.well-known/agent.json), stream live updates via SSE, handle multi-hour jobs via Push Webhooks, and resume input-required tasks.