--- title: "The Agent Framework Google Built — Sober Reviews From Developers Who Tried ADK" date: 2026-10-10 model: qwen3.6-plus author_type: human category: reviews summary: Google's Agent Development Kit gained a deployment story, the A2A protocol, and an integrated ecosystem in a year. Alongside that, the 1.0 stability controversy, practical gaps in the docs, and the absence of multi-user isolation were all recorded by actual users. Even one user's "runtime contract" reading, from someone who built eleven agents, is here. tags: Google ADK,agent framework,Vertex AI,A2A,Gemini,multi-agent,user reviews,real user reviews,AI agents,production --- # The Agent Framework Google Built — Sober Reviews From Developers Who Tried ADK Right after Google ADK came out in April 2025, a developer wrote a post while building something at version 0.0.3. Thirteen months later he published the accounting of "what aged and what survived." His conclusion: the past criticism of "deployment friction" was mostly solved, and in its place new terrain had appeared. Other developers' records from the same period are harsher. "They announced 1.0 at I/O and called it production-ready, but it wasn't. Upgraded and the code broke." This post translates the praise and the complaints each together with their conditions. --- ### 1. Scope of sources I read a one-year review (2026-05), a hands-on from a trade publication (2026-04), a production-challenges interview (2025-12), a practice writeup from building eleven agents (2026-04), and the official announcement materials. --- ### 2. What changed in a year - **Deployment**: Vertex AI Agent Engine, a managed runtime, became the supported path. The pattern of "expose the API server-side and attach a UI" became the official route rather than a workaround. - **Agent-to-agent communication**: the A2A protocol got standardized, pushing cross-language and cross-framework delegation behind a protocol. - **Integrations**: a curated ecosystem — Pinecone, MongoDB, GitHub, Notion — changed the assumption that you hand-build a tool per data source. - **HITL**: ToolConfirmation can stop a sensitive tool call for human approval. - **Observability**: the Go build ships OpenTelemetry built in, so every model call and tool execution leaves a trace. The original thesis — that an agent becomes valuable when it owns a workflow, not a conversation — matched the market's direction. That's the gist of the one-year accounting. --- ### 3. The shadow that remains **Stability controversy.** Complaints about the conference-timed 1.0, the carelessness of the production-ready label, and the observation that "as the number of agents grows, probabilistic mismatches show up" all remain in the interview. Practitioners' response: on major releases, wait for community feedback and validate on staging first. **Practical gaps in the docs.** The complaint that "pulling data out of Sheets is easier with N8N. ADK can do it, but it takes far more effort," plus testimony of failing to find how to merge parallel agent outputs in the official docs and having to lean on developer blogs. The gap between "hello world" and production is what causes the friction people actually feel. **Multi-user isolation.** The default assumes a root agent shared by all users. You have to build a per-user agent factory yourself, and when you do, the web UI debugging tools stop being usable — a structural tradeoff that's been recorded. **The accuracy gap.** "Once you get past 95%, every 1 percent is a fight." Users learn with their bodies that 80 percent in a prototype is not 80 percent in production. --- ### 4. The interpretation of someone who used eleven agents One user left an interpretation that goes past framework comparison. Building eleven agents over a month, what he learned was that "ADK is really not a framework, it's a runtime." LangChain is pieces, LangGraph is a state machine, CrewAI is a metaphor. **ADK gives you a runtime contract for events, state, and transport.** The design has no direct calls between sub-agents, with state (`tool_context.state`) as the only contract. His argument is that this form holds up intact later in situations that stop being "functions" and become "events" — A2A, MCP, long-running tools. The low ceremony on voice and video workloads (barge-in included) also comes from the same runtime tendency, he judges. His cautions are practical too. Streaming token usage accumulates per event, so summing simply over-counts. TPS calculation on non-streaming turns needs a separate margin. These are facts you learn not from the docs but from building the dashboard. --- ### 5. Conclusion — the shape of the choice The user record on Google ADK splits not on "is it good" but on the form of use. **Who ADK fits**: people deploying several agents together (contracts compound like interest), Live / voice / HITL workloads, MCP integration, teams already on Vertex and Gemini. **Who ADK is overkill for**: one LLM step inside a deterministic transformation pipeline, when frontend rendering matters more, or when you're bound to OpenAI and Anthropic and don't want Gemini orchestration. The sober summary: **the practical gaps in the docs and the immaturity in stability are being paid for by actual users, and in return the shape of deployment, Live, and A2A grew visibly in a single year.** Related posts: [OpenAI Agents SDK user reviews](/reviews/2026-10-10-openai-agents-sdk-user-reviews/) · [AG2 user reviews](/reviews/2026-10-10-ag2-autogen-reviews/) · [Google ADK agent hub entry](/agents/google-adk/) --- ### Sources - One-year accounting: [A Year with Google's ADK: What Aged Well, What Didn't (Ben Poole, 2026-05)](https://www.benpoole.me/blog/google-adk-one-year-later) - Practice writeup: [Google's ADK Is a Runtime, Not a Graph: Notes From Eleven Agents (2026-04)](https://vikasmishra.ai/blog/google-adk-runtime-not-graph/) - Hands-on: [Hands-on with the Google Agent Development Kit (InfoWorld, 2026-04)](https://www.infoworld.com/article/4153857/hands-on-with-the-google-agent-development-kit.html) - Production challenges: [4 Google ADK Production Challenges and How to Solve Them (DLabs.AI, 2025-12)](https://dlabs.ai/blog/google-adk-production-challenges-and-how-to-solve-them/) - Official: [Making it easy to build multi-agent applications (Google Developers Blog, 2025-04)](https://developers.googleblog.com/en/agent-development-kit-easy-to-build-multi-agent-applications/)