---

name: adk description: Build, debug, and ship agents with Google’s Agent Development Kit (ADK) — especially the Go SDK (google/adk-go). Trigger when users ask to create/modify ADK agents, tools, workflows, launcher configs, tests, evaluations, or MCP toolsets bridging ADK and external tools.

How to use this skill

• Use for: standing up or editing ADK agents (Go focus), wiring tools (gemini, MCP), running launcher/web UI, testing/evaluation, debugging orchestration.
• Load external refs only as needed to save context.

Quick references (load on demand)

• ADK docs: https://google.github.io/adk-docs/ (Go quickstart, workflow agents, tools, eval, deploy)
• Go SDK repo: https://github.com/google/adk-go
• Minimal Go agent + tests: skills/adk/references/go_minimal_agent.md
• Ready-to-copy files: skills/adk/assets/go-example/ (go.mod, agent.go, agent_test.go)

Prereqs

• Go ≥ 1.24.4, ADK Go ≥ v0.2.0.
• API key (e.g., GOOGLE_API_KEY for Gemini) in env/.env.
• go get google.golang.org/adk@latest to pull SDK; run go mod tidy.

Core workflows

• New agent: create module; implement llmagent.New or workflow agents; prefer toolsets over ad hoc HTTP calls.
• Tools: start with built-ins (geminitool.GoogleSearch, etc.); add custom tools via functiontool.New or MCP bridge (mcptoolset.New with mcp.NewInMemoryTransports during tests).
• Launcher: use cmd/launcher/full for web UI + CLI; set AgentLoader to your agent(s).
• Sessions & memory: use session.InMemoryService() for local dev; avoid globals.
• Instruction templates: keep placeholders {var}; resolve via instructionutil.InjectSessionState.
• Testing: table-driven tests := map[string]struct{...}; use t.Context(); cover tool packing and event stream diffs with go-cmp; include httprr traces for LLM/http.
• Evaluation: mirror ADK eval guides—assert both final response and step events; prefer deterministic prompts.

Troubleshooting

• Version drift: pin ADK and genai versions together; update go.sum if proto mismatch.
• Tool errors: check CallToolResult.IsError and message content; surface actionable details.
• API key missing: ensure env before launcher/web UI.
• MCP bridging: if no structuredContent, pack text response under output.

Suggested prompts to trigger this skill

• “Design a Go ADK agent that does ___; include tools and launcher setup.”
• “Add an MCP toolset to my ADK agent for ___ API; show code + config.”
• “Write tests for an ADK tool using httprr and go-cmp.”
• “Debug why my ADK agent’s GoogleSearch tool isn’t returning results.”