ADK (Agent Development Kit) is Google’s open-source framework for building AI agents in a way that “feels like software development” — explicit orchestration, structured context, and multi-language support, rather than a thin prompt wrapper.
Positioning: framework, not toolkit
Google has been explicit that ADK is not a grab-bag of helper functions — it’s an agent execution framework: it owns the orchestration loop, context assembly, and evaluation harness, the same way a web framework owns request routing rather than leaving you to wire it yourself.
Language support
ADK is unusually broad here compared to most agent frameworks (which are Python-only):
| Language | Status |
|---|---|
| Python | Primary |
| TypeScript/JavaScript | Supported |
| Go | Supported |
| Java | 1.0 shipped |
| Kotlin | Supported |
Orchestration model
ADK gives you two ways to control agent behavior, and expects most real systems to mix both:
Workflow agents (predictable) Dynamic routing (adaptive)
│ │
fixed pipeline steps agent-coordinated delegation
(sequential / parallel / loop) agents decide the order
It natively supports multi-agent composition — specialized agents that collaborate and delegate, conceptually similar to What is CrewAI‘s Crews, but tied to Google’s own runtime and context-management model rather than being runtime-agnostic. Cross-agent (not just cross-framework) communication runs over Google’s A2A protocol (Agent-to-Agent) — the piece CrewAI leaves to you.
v2.0 added graph-based workflows — the ability to weave deterministic code and adaptive LLM reasoning into the same execution graph, instead of choosing one mode per agent. This sits between the two poles above rather than replacing them.
Model flexibility
ADK isn’t a Gemini-only framework, despite being Google’s own tooling — it routes to whatever model backend a step needs:
| Backend | Notes |
|---|---|
| Gemini | First-party, tightest integration |
| Claude, Gemma | Supported directly |
| Ollama, vLLM, LiteLLM, LiteRT-LM | Local / self-hosted model routing |
| Vertex AI Agent Engine, Apigee AI Gateway | Managed / gateway-fronted deployment |
This matters for the same reason OTel-native instrumentation matters over vendor SDKs: the orchestration layer doesn’t lock you into one model vendor even though the managed runtime (Agent Engine) is Google Cloud-only.
Evaluation framework
Evaluation is a first-class ADK component, not an afterthought bolted on post-launch: criteria-based scoring, user simulation, and environment simulation, with custom metrics for domain-specific correctness checks. This is the harness ADK expects you to run before promoting an agent out of dev.
Context management — the actual differentiator
Most agent frameworks handle context by concatenating strings into the prompt until the window overflows. ADK treats context closer to source code: every token in the context window has to earn its place.
Session state + Memory + Tool outputs + Artifacts
│
▼
ADK context assembler:
- filters irrelevant events
- summarizes older turns
- lazy-loads artifacts (not eagerly inlined)
- tracks token budget continuously
│
▼
Structured context ──▶ LLM call
This is the same problem What is Mem0 solves for memory specifically — ADK’s version is broader, covering session state and tool artifacts too, not just long-term memory facts.
Deployment targets
adk deploy ──▶ Vertex AI Agent Engine (see [[vertex-ai]]), now under [[gemini-enterprise-agent-platform]]
├──▶ Google Cloud Run
├──▶ Google Kubernetes Engine (GKE)
└──▶ any other container/Kubernetes environment
ADK code isn’t locked to Google Cloud at deploy time — the framework is open source and the agent can run anywhere containers run — but the managed runtime option (Agent Engine) is Google Cloud only.
Where it fits
| Concern | ADK | What is CrewAI |
|---|---|---|
| Languages | Python, TS/JS, Go, Java, Kotlin | Python only |
| Model backend | Gemini, Claude, Gemma, Ollama, vLLM, LiteLLM | OpenAI/Anthropic + Snowflake Cortex |
| Context management | Built-in, structured | Left to you / pluggable |
| Managed runtime | Vertex AI Agent Engine (optional) | None — bring your own |
| Database tool access | First-class integration with What is MCP Toolbox | Wrap as a generic tool |
Why it’s on the backlog: if the h-aiops SRE-agent line ever needs multi-language agents (a Go-based collector talking to a Python reasoning layer) or a managed deploy target, ADK is the framework built for exactly that, at the cost of coupling more tightly to Google Cloud’s runtime.
Local graph
Linked from 11 notes
What is CrewAI
Open-source Python framework for multi-agent orchestration — role-based 'Crews' for autonomous collaboration and event-driven 'Flows' for precise control, now a de facto standard for production agentic pipelines.
What is Harness Engineering
The discipline of designing everything that wraps a raw LLM into a reliable agent — tool contracts, system prompt architecture, context/memory management, permission gates, and feedback loops. Distinct from prompt engineering and model training.
What is Gemini Enterprise Agent Platform
Google's April 2026 unification of agent tooling — a four-stage lifecycle (build, scale, govern, optimize) wrapping Agent Studio/ADK, a stateful Agent Runtime, an Identity/Registry/Gateway governance stack, and native A2A + MCP interop.
What is MCP Toolbox
Google's open-source MCP server for databases (formerly Gen AI Toolbox for Databases) — production-ready (v1.0.0) prebuilt tools that connect agents and IDEs directly to AlloyDB, Spanner, Cloud SQL, BigQuery, and more, with OAuth2 zero-trust gating.
What is Mem0
Universal memory layer for AI agents — combines vector search, a knowledge graph, and key-value caching behind one API, so any framework can bolt on persistent, cross-session memory in under a day.
What is OpenClaw
Self-hosted, model-agnostic personal AI agent (by Peter Steinberger) that gets full computer access — browser, filesystem, shell — and is reachable from 29+ chat platforms, positioned against SaaS agent walled gardens.
What is Vertex AI
Google Cloud's managed ML/AI platform — as of 2026 rebranded and consolidated into the Gemini Enterprise Agent Platform, bundling 200+ foundation models, Agent Builder, and a managed agent runtime (formerly 'Agent Engine').
Agentic AI: Projects & Engineering Mastery
A book-shaped table of contents for Agentic AI: Projects & Engineering Mastery: hands-on practitioner builds, Principal/Staff-level technical leadership, and the lookup appendices and vendor/framework reference notes for the whole series. Book 6 of the AI Systems Engineering series.
What is Hermes Agent
Nous Research's open-source, self-hosted AI agent that runs continuously on your own server, builds persistent memory over time, and distills its own reusable 'skills' from completed tasks.
7. Google ADK
Covers Google's Agent Development Kit — its workflow/dynamic-routing composition model, native tool-integration story, and the deployment path onto Vertex AI Agent Engine that is the actual site of vendor coupling, not the framework code itself.
Building & Evaluating Agents
A book-shaped table of contents for Building & Evaluating Agents: the architectural core of agent design — building single-agent systems, multi-agent systems, evaluation, and the agent framework landscape. Book 3 of the AI Systems Engineering series.
Related notes
What is Gemini Enterprise Agent Platform
Google's April 2026 unification of agent tooling — a four-stage lifecycle (build, scale, govern, optimize) wrapping Agent Studio/ADK, a stateful Agent Runtime, an Identity/Registry/Gateway governance stack, and native A2A + MCP interop.
What is CrewAI
Open-source Python framework for multi-agent orchestration — role-based 'Crews' for autonomous collaboration and event-driven 'Flows' for precise control, now a de facto standard for production agentic pipelines.
What is Hermes Agent
Nous Research's open-source, self-hosted AI agent that runs continuously on your own server, builds persistent memory over time, and distills its own reusable 'skills' from completed tasks.
What is OpenClaw
Self-hosted, model-agnostic personal AI agent (by Peter Steinberger) that gets full computer access — browser, filesystem, shell — and is reachable from 29+ chat platforms, positioned against SaaS agent walled gardens.