ADK (Agent Development Kit)

google‘s framework for building AI agents adk-agents-with-skills. Its relevance to this wiki is that it adopts the open agentskills-spec: ADK agents “load specialized domain expertise on demand rather than embedding everything in a monolithic system prompt,” via agent-skills with progressive disclosure.

Skill-authoring patterns it supports

A progression of four: inline skills → file-based skills (with reference material) → external imports → self-generating meta-skills (an agent that writes its own skills). The guide advises reviewing AI-generated skills before deployment, like code review.

Significance

ADK is this wiki’s evidence that the agent-skills format is cross-vendor, not Anthropic- specific: it sits alongside Claude Code / Cursor / Gemini CLI as one of “40+” products consuming the same skill files. Maker google thus joins anthropic in the agent- substrate picture. See agent-skills for the broader pattern.

Beyond skills, ADK also provides durable-agents infrastructure — durable state machines, persistent sessions (DatabaseSessionService), and webhook-driven pause/resume for multi-week, human-in-the-loop workflows (adk-long-running-agents). So ADK spans both the skills and the durable-execution sides of an agent framework.

ADK Go 2.0 — the graph-based workflow engine

The 2.0 release (30 Jun 2026, adk-go-2) rebuilds the framework around a graph of typed nodes and conditional edges run by a scheduler — its claim: “a graph is an agent,” so the same runtime drives a single LlmAgent and a full multi-agent workflow. It makes human-in-the-loop first-class (any node pauses into session state and resumes by handoff or re-entry, durable across process restarts) and moves reliability into the runtime (per-node retries/timeouts, WithMaxConcurrency, branch isolation). LLM agents gain Chat/Task/ SingleTurn modes; ToolContext+CallbackContext merge into one agent.Context. The resume format is shared with Python ADK 2.0, so a paused workflow can resume across runtimes. This is the deepest expression of ADK’s durable-execution side and folds orchestration into the base runtime.

Multi-agent orchestration over A2A

ADK also supplies the orchestration primitives for cross-vendor, cross-language teams: a SequentialAgent chains sub-agents, and RemoteA2aAgent wraps any remote A2A endpoint as a local sub-agent (no custom HTTP client), with ToolContext.state as the shared checkpoint store. adk-a2a-cross-language shows it driving a Go validator from a Python orchestrator — so ADK is the orchestrator that turns A2A’s agent↔agent wire format into a running multi-language pipeline.

In production — the AI Race Coach

ai-race-coach (Google Developers Blog case study) shows ADK orchestrating “a collection of agents” in a live real-time system — a driving coach at Sonoma Raceway — rather than in a docs example. It ran alongside Antigravity (stateful telemetry orchestration) and a hybrid model split: Gemma 4 local for zero-latency audio alerts, the Gemini API for cloud post-session modelling. Evidence the ADK orchestration primitives ship into the field, not just the reference guides. (T3 vendor case study.)

google · agentskills-spec · agent-skills · adk-agents-with-skills · durable-agents · adk-long-running-agents · a2a-protocol · adk-a2a-cross-language · agent-orchestration · adk-go-2 · graph-based-agent-workflow · ai-race-coach