Key takeaways
- The Python repo passed 20,000 GitHub stars in just over a year (launched April 2025), shipping on a weekly release cadence with v1.35.0 landing June 10, 2026 alongside a parallel v2.x line that adds a graph-based workflow runtime
- Hybrid architecture is the differentiator — deterministic workflow agents (sequential, parallel, loop) compose with LLM-routed dynamic delegation in hierarchical agent teams, connected across vendors via the A2A protocol
- Gemini-optimized but model-agnostic via Vertex AI Model Garden and LiteLLM (Anthropic, Meta, Mistral, and more), with managed enterprise deployment to Vertex AI Agent Engine, GKE, or Cloud Run
FAQ
What is Google ADK?
Google ADK (Agent Development Kit) is an open-source, code-first framework for building, evaluating, and deploying AI agents, with SDKs in Python, TypeScript, Go, Java, and Kotlin.
How much does Google ADK cost?
The framework is free and Apache 2.0 licensed. Managed deployment via Vertex AI Agent Engine, Cloud Run, or GKE is billed at standard usage-based Google Cloud rates; ADK itself has no separate price.
What models does Google ADK support?
ADK is optimized for Gemini but works with any model in Vertex AI Model Garden and, via LiteLLM, with third-party providers including Anthropic, Meta, Mistral AI, and AI21 Labs, plus local serving through Ollama and vLLM.
How is Google ADK different from LangChain?
LangChain is a vendor-neutral ecosystem with 1000+ integrations; ADK is Google's opinionated framework that pairs deterministic workflow agents with LLM-driven delegation and offers the tightest path to Google Cloud managed agent infrastructure.
Executive Summary
Google ADK (Agent Development Kit) is Google's open-source, code-first framework for building, evaluating, and deploying AI agents, announced at Google Cloud NEXT on April 9, 2025 — the same framework that powers agents inside Google products like Agentspace and the Customer Engagement Suite.[1] The Python repo stands at 20,000+ GitHub stars with 3,500+ forks as of June 2026, under the Apache 2.0 license.[2] Its core bet is a hybrid architecture: deterministic workflow agents (sequential, parallel, loop) compose with LLM-routed dynamic delegation in hierarchical multi-agent teams, with the A2A protocol for cross-vendor agent interoperability.[3]
ADK is Gemini-optimized but deliberately multi-model — any model in Vertex AI Model Garden works, and LiteLLM integration adds Anthropic, Meta, Mistral AI, and AI21 Labs.[1] The enterprise story is the strongest in its class: managed sessions, Memory Bank, observability, and evaluation via Vertex AI Agent Engine, plus GKE and Cloud Run deploy paths.[4]
| Attribute | Value |
|---|---|
| Company | Google (Google Cloud / DeepMind ecosystem) |
| Launched | April 9, 2025 (Google Cloud NEXT)[1] |
| Funding | N/A — Google product |
| GitHub Stars | 20,000+ (Python repo, as of June 2026)[2] |
| License | Apache 2.0[2] |
Product Overview
ADK is a code-first toolkit: agents are defined in ordinary code, debugged in a built-in web dev UI, evaluated with an integrated framework, and deployed as containers or to managed Google Cloud runtimes.[1] Beyond the flagship Python SDK, Google maintains TypeScript, Go, Java, and Kotlin implementations — the Go SDK alone has 8,100+ stars as of June 2026.[3][5]
Key Capabilities
| Capability | Description |
|---|---|
| Workflow agents | Deterministic sequential, parallel, and loop pipelines[3] |
| LLM-routed delegation | Dynamic routing where a model decides which sub-agent handles a task |
| Hierarchical multi-agent | Specialized agent teams that collaborate and delegate[4] |
| A2A protocol | Agents expose and consume capabilities across deployments and vendors[3] |
| Tool ecosystem | Function tools, OpenAPI tools, and MCP (Model Context Protocol) tools[3] |
| Dev UI + evals | Visual web UI for debugging; criteria-based evaluation, user and environment simulation[3] |
| Streaming | Bidirectional audio and video streaming for natural interactions[1] |
Product Surfaces / Editions
| Surface | Description | Availability |
|---|---|---|
| ADK SDKs | Python, TypeScript, Go, Java, Kotlin frameworks (Apache 2.0)[3] | GA |
| ADK web UI | Local visual dev/debug interface | GA |
| Vertex AI Agent Engine | Managed runtime with sessions, Memory Bank, observability[4] | GA |
| Cloud Run / GKE | Containerized self-managed deployment[4] | GA |
Technical Architecture
ADK separates deterministic orchestration from autonomous reasoning: workflow agents handle predictable pipelines while LLM agents handle dynamic routing and delegation, and both compose into hierarchical trees.[3]
Language: Python 3.11+ for the flagship SDK[2]
pip install google-adk
Key Technical Details
| Aspect | Detail |
|---|---|
| Deployment | Vertex AI Agent Engine (managed), Cloud Run, GKE (Kubernetes), containers[4] |
| Model(s) | Gemini (optimized), Vertex AI Model Garden; Anthropic, Meta, Mistral AI, AI21 via LiteLLM; Ollama/vLLM local[1][3] |
| Integrations | MCP tools, OpenAPI tools, function tools, A2A protocol[3] |
| Observability | OpenTelemetry-compatible tracing, logging, monitoring; evaluation (offline, simulated, continuous online)[4] |
| Open Source | Yes (Apache 2.0)[2] |
Release Cadence
The Python repo has shipped 65 releases on a near-weekly cadence since launch, currently across two lines: the stable 1.x series (v1.35.0, June 10, 2026) and a 2.x series (v2.2.0, June 4, 2026) introducing a graph-based workflow runtime and a task API for agent-to-agent delegation with routing, loops, retries, and human-in-the-loop.[6][2]
Strengths
- Hybrid orchestration model — Deterministic workflow agents plus LLM-routed delegation in one framework, avoiding the "all autonomy or all DAG" tradeoff[3]
- Strongest managed-runtime story — Vertex AI Agent Engine provides sessions, Memory Bank, observability, evaluation, and an Agent Gateway out of the box[4]
- True multi-language — Five official SDKs (Python, TypeScript, Go, Java, Kotlin), rare among agent frameworks[3]
- Battle-tested internally — The same framework powers Google Agentspace and the Customer Engagement Suite[1]
- Open standards posture — A2A for agent interoperability and MCP for tools, rather than proprietary protocols[3]
- Rapid, sustained velocity — 65 releases in ~14 months, with v2.x landing major runtime upgrades[6]
- Genuinely multi-model — LiteLLM and Model Garden integration despite the Gemini-first framing[1]
Cautions
- Google Cloud gravity — The framework is portable, but the managed sessions/memory/observability tier lives in Vertex AI Agent Engine; full value assumes GCP[4]
- Gemini-first optimization — Multi-model works, but the tightest integration and defaults favor Gemini[1]
- Dual release lines — Parallel 1.x and 2.x series shipping simultaneously (v1.35.0 and v2.2.0 within a week of each other) creates upgrade-path ambiguity[6]
- Young ecosystem — Launched April 2025; community content, third-party integrations, and production case studies trail LangChain and CrewAI[1]
- Uneven SDK maturity — Python (20,000+ stars) and Go (8,100+) are healthy; the Kotlin SDK is nascent[2][5]
- Google product-risk reputation — Developers reasonably price in Google's history of sunsetting products, though ADK's role inside Agentspace argues for staying power
Pricing & Licensing
| Tier | Price | Includes |
|---|---|---|
| ADK (all SDKs) | Free | Full framework, dev UI, evals (Apache 2.0)[2] |
| Vertex AI Agent Engine | Usage-based GCP pricing | Managed runtime, sessions, Memory Bank, observability[4] |
| Cloud Run / GKE | Usage-based GCP pricing | Self-managed container deployment[4] |
Licensing model: Open source (Apache 2.0) framework; commercial monetization happens through Google Cloud consumption, not the framework itself.[2]
Hidden costs: Agent Engine pricing is not listed on the ADK docs — it routes to Agent Platform and Generative AI pricing pages, so production costs are workload-dependent (compute, model tokens, session/memory storage). LLM provider costs are separate when using third-party models via LiteLLM.[4]
Competitive Positioning
Direct Competitors
| Competitor | Differentiation |
|---|---|
| LangChain/LangGraph | LangChain has the largest ecosystem (1000+ integrations) and vendor neutrality; ADK has tighter GCP-managed infrastructure and five official SDKs |
| Microsoft Agent Framework | The Azure-ecosystem mirror image of ADK; choose based on which cloud you live in |
| OpenAI Agents SDK | Lighter-weight and OpenAI-centric; ADK offers richer deterministic workflow primitives and multi-cloud model access |
| CrewAI | CrewAI leads in role-based multi-agent teams and Fortune 500 adoption claims; ADK leads in language coverage and managed runtime depth |
When to Choose Google ADK Over Alternatives
- Choose ADK when: You run on Google Cloud, want Gemini-optimized agents with a managed runtime, or need deterministic workflows and LLM delegation in the same framework
- Choose LangChain when: You want maximum integrations, vendor neutrality, and the largest community
- Choose Microsoft Agent Framework when: Your organization is Azure/.NET-centric
- Choose OpenAI Agents SDK when: You're committed to OpenAI models and want the thinnest abstraction layer
Ideal Customer Profile
Best fit:
- Google Cloud-native organizations standardizing on Vertex AI
- Teams that need both deterministic pipelines and autonomous agent delegation
- Polyglot engineering orgs (Python, TypeScript, Go, Java teams sharing one framework)
- Enterprises that want managed sessions, memory, and observability without building them
- Teams betting on open protocols (A2A, MCP) for agent interoperability
Poor fit:
- AWS- or Azure-committed organizations
- Teams that want a vendor-neutral framework with no cloud gravity
- Projects needing the deepest third-party integration catalog today
- Developers wary of Google's product-sunset track record
Viability Assessment
| Factor | Assessment |
|---|---|
| Financial Health | Backed by Google — no funding risk; monetized via GCP consumption |
| Market Position | Fast challenger — 20,000+ stars in ~14 months, behind LangChain/CrewAI in mindshare[2] |
| Innovation Pace | Rapid — 65 releases since April 2025, weekly cadence, v2.x workflow runtime landing[6] |
| Community/Ecosystem | Growing — five official SDKs, 3,500+ forks on Python repo, strong samples repo[2] |
| Long-term Outlook | Positive — anchors Google's agent strategy and powers internal products[1] |
ADK's viability is really a question about Google's agent strategy, and the signals are strong: it powers Agentspace, it anchors the Vertex AI Agent Builder story, and it ships weekly across five languages. The risk is not abandonment but churn — the dual 1.x/2.x release lines show a framework still settling its core runtime abstractions.
Bottom Line
Google ADK is the most credible cloud-vendor agent framework: genuinely open source, genuinely multi-model, and paired with the deepest managed-runtime offering of any framework in this category. The hybrid workflow-plus-delegation architecture is the right design for production systems that need predictability in some paths and autonomy in others.
Recommended for: GCP-native enterprises building production agents who want managed sessions, memory, and observability, and polyglot teams that need one framework across Python, TypeScript, Go, and Java.
Not recommended for: Vendor-neutral purists, AWS/Azure-committed shops, or teams that need today's largest integration ecosystem and community knowledge base.
Outlook: With 20,000+ stars in its first 14 months and weekly releases, ADK is closing the gap with incumbent frameworks faster than any entrant in the category.[2][6] The v2.x graph-based workflow runtime and the A2A protocol bet position it as infrastructure for cross-vendor agent ecosystems, not just GCP workloads — watch whether the 2.x line consolidates the community or fragments it.
Research by Ry Walker Research • methodology