Google ADK

ADK is a code-first, open-source toolkit for building, evaluating and deploying agents and multi-agent systems, built around an event-driven runtime and hierarchical "agent trees" (agents composed of sub-agents) rather than the graph-of-nodes model LangGraph uses or the role-based crews CrewAI uses. It ships the Agent2Agent (A2A) protocol for inter-agent communication out of the box, and reviewers who've used both note ADK asks more of you upfront to learn the model but scales more cleanly into production once you have (google/adk-python GitHub description; docs at adk.dev; ZenML and 1337skills framework comparisons).

google/adk-python has 21,206 stars, 3,868 forks, 537 open issues and roughly 400+ contributors as of 2026-08-21, created April 2025. It was pushed to within the last day and ships on a fast, structured cadence: the current PyPI release is 2.7.1 (uploaded 2026-08-17), and Google maintains a parallel v1.x line that still gets backport releases (v1.39.0 shipped the same day as v2.7.1). Release notes are detailed conventional-commit changelogs with explicit "Breaking changes" callouts, which is more disciplined than most frameworks this size.

Beyond the flagship Python package, Google maintains adk-go (8,678 stars), adk-java (1,690 stars, Apache 2.0, also pushed within the last day) and adk-js for TypeScript (1,356 stars, published to npm as @google/adk, currently v1.6.0), plus a Kotlin SDK referenced in the docs. All the repos checked show commits within the last 24 hours, which is a stronger signal than a framework that ports its Python API to other languages once and lets them drift.

ADK itself is free and Apache-2.0 licensed — no field-of-use restriction. The framework's actual production pitch is the deployment path: agents built with ADK deploy to Cloud Run, GKE, or Vertex AI's managed "Agent Engine" runtime (recently rebranded, alongside the rest of Vertex AI's agent tooling, into Google's "Gemini Enterprise Agent Platform"). That managed runtime bills separately from ADK: $0.085 per vCPU-hour and $0.009 per GiB-hour, with a monthly free tier of 50 vCPU-hours and 100 GiB-hours per account. On top of that you pay for whatever model you call — Gemini or otherwise — which is typically the larger cost.

In April 2026, GitHub/NVD published CVE-2026-4810 (GHSA-rg7c-g689-fr3x), a CRITICAL-severity code injection and missing-authentication flaw allowing unauthenticated remote code execution against ADK instances (Python OSS, Cloud Run and GKE) running versions 1.7.0 through 1.28.1 and 2.0.0a1–2.0.0a2. It was patched in 1.28.1 and 2.0.0a2 — anyone still on an affected version needs to redeploy, and if you run adk web locally you need to upgrade that too. Separately, in June–July 2026, researchers at Pillar Security disclosed that Google's own CI/CD automation for the adk-python repo — three ADK-built GitHub Actions workflows (issue-analyze.yml, issue-fix.yml, pr-analyze.yml) — could be manipulated via a crafted public GitHub issue to impersonate the trusted adk-bot account and trigger a privileged code-fixing agent, achieving code execution on CI runners. Google removed the three workflows and had a fix in place by 2026-07-21. That second incident is not a flaw in the shipped ADK package — it's Google's own dogfooding getting hit — but it's a concrete, dated example of the agent-to-agent trust-boundary problem any multi-agent system built on ADK (or anything like it) has to design around.

Independent comparisons consistently note ADK has a smaller third-party plugin/integration ecosystem than LangChain/LangGraph or CrewAI, reflecting its shorter track record (first released April 2025) and its natural gravity toward the Google Cloud/Gemini stack. If you need to mix models and providers freely with a mature persistence and human-in-the-loop layer, or want the fastest path to a simple role-based agent crew, reviewers point to LangGraph and CrewAI respectively as the better fit; ADK's advantage shows up specifically when you're standardizing on Gemini and want an integrated path to Vertex AI deployment.