Skip to content

WHY GOOGLE WILL FAIL IF YOU DO NOT CHANGE COURSE #7951

Closed

Description

@Tech4Biz

From: Gregory — Builder, Product Owner, and P&L Leader (Telecom & Finance)
To: Google Chief Product Officer, VP of Product (Vertex AI / Google Cloud), and Google ADK Leadership
Re: You Are Losing the Developer & Enterprise Market Because Ivory-Tower PMs Are Shipping Science Projects Instead of Products

  1. The Disconnect: World-Class Research, Amateur Product Ownership
    Google has the best foundational AI research on the planet. Yet in the trenches—where actual builders are shipping real production applications on Cloud Run, Vertex AI, and Google ADK (google-adk 2.3.0, gemini-3.5-flash)—your product experience is broken.
    I am building and running a multi-application Chief of Staff ecosystem (Tish-COS + ManageMyLife v2) on your stack. Instead of your framework accelerating me, I have been forced to act as an unpaid systems integrator—hand-rolling custom Firestore session rotation, custom cosine-vector memory services, custom OIDC plumbing, and manual Python tool wrappers for capabilities that Google already built (OpenAPIToolset, VertexAiSessionService, VertexAiMemoryBankService, Grounding, and user: state persistence) but buried under fragmented documentation and unopinionated primitives.
    Even worse: when we use AI coding assistants (including Gemini itself!) to build on Google ADK, your own models don't even default to using your native managed primitives—they write brittle custom workarounds from scratch. That is a catastrophic Developer Experience (DX) failure.
  2. Why Google Will Lose the Market If You Do Not Course-Correct
    In Telecom and Finance—industries where I have owned products and managed real P&Ls—if you ship raw engine parts to a customer and tell them to weld their own chassis together, your churn hits 100% and a competitor eats your lunch.
    Builders do not want a box of 50 disconnected APIs. They want opinionated, production-ready reference capabilities (like a plug-and-play Chief of Staff / Executive Agent) that work out of the box.
    An AI agent in 2026 should not behave like a 2015 Jira bot. When an agent built on Google ADK encounters a missing parameter or capability, it should not dump a backlog ticket onto the human user's lap. It should natively inspect the backend OpenAPI spec, persist standing executive rules deterministically, and use Google Search Grounding + ADK documentation to research, vet, and propose its own architectural upgrade.
    You are forcing your champions to do the job of your Product Managers. Every hour a builder spends diagnosing why EventsCompactionConfig or session rotation wiped out their UI chat history is an hour they are pushed toward a competitor that abstracts the plumbing away.
  3. The Root Cause: You Need P&L Owners & Builders at the Helm
    Your current product gaps are the direct result of product managers who manage slide decks and feature checkboxes rather than living inside the product they ship and owning a real commercial P&L.
    Someone who has actually run a P&L across high-stakes verticals like Telecom and Finance—and who sits down at the keyboard to build production systems with these tools every single day—would have caught these friction points in week one:
    They would have shipped a native, zero-boilerplate Chief of Staff (CosAgent) blueprint in ADK Day 1.
    They would have made OpenAPI auto-ingestion (OpenAPIToolset) and managed memory/session services the default path, not an obscure footnote.
    They would have separated UI Transcript History from LLM Context Window Compaction at the framework level so developers never have to choose between token cost and seeing their own conversation history.
  4. My Challenge to Google Leadership
    If you are serious about winning the agentic AI platform war instead of watching developers abandon Vertex and ADK out of sheer integration fatigue, stop hiring career feature-managers and hire builders who have carried a real P&L.
    I am living inside your stack every day, pushing gemini-3.5-flash and google-adk further in production than the people writing your specs. Look at the architectural session logs attached to this feedback. Hire someone like me to run Product for Agentic Developer Experience, fix the product from the outside-in, and stop shooting yourselves in the foot.

Activity

  1. added
    question[Component] This issue is asking a question or clarification
    on Nov 3, 2026
  2. kmaphoenix commented on Nov 4, 2026

    @kmaphoenix

    @Tech4Biz

    Hey Gregory!
    Thanks for the candid feedback.

    I'm the new PM for ADK and took over in the last couple of months.
    I don't have all the answers yet, but we're making some significant changes to ADK based on the feedback we've been hearing from our customers and community.

    I'm a firm believer that code wins arguments, and that hands-on building is the best way to form strong product opinions.
    I build and use ADK agents every day in my own work, so feedback like this resonates with me.

    A few things to know:

    1. Agents CLI. If you haven't checked it out yet, you should. It's the best way to scaffold, build, eval, and deploy ADK agents on GCP. It also gets at your point about coding agents writing workarounds from scratch. The skills that ship with it teach your coding agent the ADK and GCP best practices across the whole Agent Development Lifecycle, so it reaches for the built-in services instead of reinventing them.
    2. adk-recipes is the new home for our reference architectures. It used to be adk-samples, but many of those samples were code snippets or had aged out. Recipes are meant to be E2E references, infra scaffolding included, that you can use as a launch point for what you're trying to build.
    3. Long Horizon is a recipe in that repo. It's an opinionated reference for building a harness based agent on ADK that can complete long horizon tasks, with cross session memory through Memory Bank, a per-user sandbox, tool guardrails, sub-agents, and a self-improvement loop. I think it's the closest match to your Chief of Staff use case. The architecture doc walks through the design.
    4. File issues for the friction points. OpenAPIToolset ergonomics, session and memory service defaults, and keeping the UI transcript separate from context compaction are all worth their own issue on adk-python. Building a framework that meets every user's needs 100% of the time is the eternal struggle. Our DevRel team is always hunting for gaps, but when you or anyone in the community hits something that should be fixed now, an issue is the fastest way to get it in front of the team so we can fix it or give guidance.

    Finally, we'd love for you to join our ADK Community Calls to stay current on announcements and what we're working on, including much of the above.
    Previous calls are recorded here.

    Thanks again for the feedback.
    I'm looking forward to digging into the issues you file, and hope to meet you on a community call soon.

    - Patrick

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Labels

question[Component] This issue is asking a question or clarification

Type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions