$ cat wiki/concepts/google-adk.md
Google ADK (Agent Development Kit)
Definition
Google's Agent Development Kit (ADK) is an open-source, code-first toolkit for building, evaluating, and deploying AI agents. As of June 2026, ADK 2.0 reached General Availability with two new capabilities: graph workflows and collaborative multi-agent systems.
The Agents CLI is a companion tool providing end-to-end agent lifecycle management (scaffolding → evals → deploy → observability → publishing) in a single interface.
Why It Matters
- Scores 2.34 (agents 1.5 × Google 1.3 × official blog 1.2) — the highest agent-tooling signal since Microsoft Build 2026
- Directly addresses the gap Karpathy identified: 89% of teams have observability but only 52% have evals — the Agents CLI packages eval capabilities as first-class
- Universal compatibility: Agents CLI works with Claude Code, Cursor, Gemini CLI, and Google's Antigravity — making it model-agnostic infra
- Multi-language (Python, Go, Java, Kotlin) lowers the barrier for mobile and backend agent coordination
State of the Art (2026-06-30)
ADK 2.0 GA Features
Graph workflows — deterministic execution composing:
- Routing and conditional branching
- Fan-out/fan-in (parallel agent dispatch and merge)
- Retry and state management
- Dynamic nodes and human-in-the-loop
- Nested workflows
Collaborative multi-agent systems — Task API for structured agent-to-agent delegation. A meta-agent assigns tasks to sub-agents via the Task API; sub-agents report completion with structured output.
Agents CLI Features
One unified tool covering the full agent lifecycle:
adk scaffold— project setup with ADK best practices- Eval integration — direct evaluation loop baked into CLI
adk deploy— deployment to Gemini Enterprise Agent Platform / Vertex AI- Observability — tracing pipeline included
- Publishing — enterprise agent registry
One setup command injects 7 ADK-specific skills into a coding agent (Claude Code, Cursor, etc.), enabling agents to build and deploy other agents.
Language Support
| Language | Status |
|---|---|
| Python | GA |
| Go | GA (ADK Go 1.0 also released June 2026) |
| Java | GA |
| Kotlin | Beta (new — mobile on-device agent coordination) |
Open Problems
- Eval standardization — ADK eval tooling covers ADK-native agents but not heterogeneous agent stacks
- Cross-platform interop — ADK's Task API is Google-native; no direct bridge to Anthropic Managed Agents or OpenAI Deployment Company APIs
- Long-horizon state management — graph workflows enable it structurally, but semantic coherence over 100+ steps remains unsolved
Key Papers
- The agents-as-building-blocks vision is consistent with Software 3.0 (Karpathy's agentic engineering paradigm)
- Scaling the Horizon, Not the Parameters: Reaching Trillion-Parameter Performance with a 35B Agent — Agents-A1: horizon scaling over parameter scaling, complementary to ADK's long-horizon trajectory infrastructure
Related Concepts
- Agents (LLM Agents) — LLM agents broadly; ADK is the tooling layer
- Claude Managed Agents — Anthropic's equivalent agent execution platform
- Google DeepMind — parent org