go-adk（去 adk）

go-adk

Expert guidance for developing agents with google.golang.org/adk (Go ADK).

When to trigger

• The user mentions Go ADK, google.golang.org/adk, adk-go, or asks how to build agents in Go.
• The task involves ADK concepts like llmagent, tools, callbacks, sessions/state/memory, workflow agents, streaming, MCP, A2A, ADK web/API runtime, launchers, deployment, evaluation, or safety controls.

Core rules

• Treat references/llms-index.md as the required ADK coverage map.
• Prefer official ADK Go APIs and examples over ad-hoc wrappers.
• Follow existing repository patterns first (imports, launcher style, layout, env handling).
• For version-sensitive behavior, check release notes and API reference before giving definitive guidance.
• Prioritize Go-specific ADK docs first; if a feature is documented only in another SDK doc, call that out explicitly.
• Keep agent names unique and never use user as an agent name.
• Use concise Description fields so delegation and tool selection remain reliable.
• Treat ADK Web as development-only unless the user explicitly asks for production runtime choices.

Workflow

1. Map task to ADK domain:

- Build, run, deploy, evaluate, safety, components, protocols, or reference. - Load only the relevant sections from references/llms-index.md.

1. Confirm execution style:

- Launcher-first app (cmd/launcher, cmd/launcher/full) or embedded runtime (runner.Run loop). - Local dev UI, CLI-only, or REST/API server flow.

1. Build the first working agent:

- llmagent.New(llmagent.Config{...}) with a concrete model (commonly gemini.NewModel(...)). - Add minimum viable Instruction, Description, and Tools.

1. Add tools safely:

- Use functiontool.New(...) for typed custom tools. - Use geminitool.GoogleSearch{} or other ADK/native integrations only when needed. - For sensitive tools, apply confirmations, auth, and policy callbacks.

1. Wire stateful execution:

- Sessions via session.Service (often session.NewInMemoryService() for local dev). - Add memory.Service only when cross-session retrieval is required. - Add state/context/artifacts management only when required by the workflow.

1. Apply control points:

- Use callbacks (Before/After Model, Before/After Tool, agent callbacks) for logging, caching, guardrails, and overrides. - Prefer plugins for reusable security policy enforcement.

1. Scale composition:

- Use workflow agents for deterministic orchestration (sequentialagent, parallelagent, loopagent). - Use multi-agent delegation (descriptions + transfer behavior) only when specialization is clear.

1. Select runtime/deployment posture:

- Local dev: CLI/Web/API. - Production: deployment target + observability + safety posture.

1. Verify behavior end-to-end:

- Validate happy-path conversation, tool errors, callbacks, session resume/rewind behavior, and evaluation criteria.

Output expectations

• Provide exact Go imports, runnable snippets, and command lines that match the chosen runtime mode.
• Call out version-sensitive behavior if the user targets older ADK tags.
• Mention the relevant ADK doc section when guidance comes from llms-index categories beyond core Go quickstart pages.
• Prefer small, testable incremental changes over large rewrites.

References

• Quickstart and launcher baseline: references/quickstart.md
• Core APIs and agent config: references/core-primitives.md
• Tools, callbacks, and guardrails: references/tools-and-guardrails.md
• Runtime modes, examples, and A2A: references/runtime-and-examples.md
• Required ADK source map (llms.txt mirror): references/llms-index.md