Bigger list today as I try to clear out the queue. Too many great pieces lately! I’m still starting a dozen open tabs I need to read over the weekend. Could be worse.
[blog] React Native solved a problem nobody had. Use Flutter. Fundamental points being made here. If you want native, don’t use a proxy for native. But if you want a rendering engine for owned experience, that’s a case for Flutter.
[blog] Introducing the Server Side Cloud Swift SDK. Swift isn’t just for UI work on Apple client devices? I’m not sure I knew that. We just shipped an SDK for those who want Swift’s concurrency benefits on the server.
[blog] Why redesign doesn’t happen (enough). Sounds like a realistic story to me. What can remove the fear of making changes, breaking other PRs, and getting stuck in merge hell?
I’m at the airport waiting to fly home after a day trip up the Bay Area. After a few hours with a fantastic audience at the local DevFest, I spent the late afternoon with customer execs who wanted to talk about software engineering with AI.
[blog] The Code Nobody Reads. Big piece, worth reading. Code review is changing, but the key will be knowing what’s worth reading and reviewing.
[article] Maintaining quality while agents write the code. Somewhat related to the first item here. We’re going to need to zoom in on the decisions that require thoughtful instructions and reviews.
[blog] Coding is NOT solved. Maybe a counter-point to other items listed here. We’re responsible for what we produce, whether we understand it or not. I don’t agree with all (most?) of this, but I’m glad to read the perspective.
[blog] Empower your agents with the Google Cloud CLI remote MCP server. CLI-over-MCP is an important pattern, especially when dealing with AI clients that won’t be allowed to call local scripts or CLIs to do stuff. I used this in my blog post from yesterday.
Last night, I went to the San Diego Padres playoff game against the Chicago Cubs. It was the rowdiest, loudest, most fun baseball game I’ve attended in my life. The downside? I’m very hoarse and had two major speaking gigs today. So yeah, I plan well.
[blog] State of agent skills. Vercel sits on a lot of data here thanks to running skills.sh. Read this for a look at what skills people are installing, for which industries, and how often.
[blog] State of Markets II. We’re getting a bunch of “state of XX” things lately. The a16z folks call out a bunch of data about companies using AI, and more.
[blog] Voice Agents Can Just Do Things. I buy this more than I did a year ago. But I also don’t want this to be the ONLY interface. Sometimes it’s better to type things out, or I’m in a space where I don’t want to be barking commands to a robot.
I have 334 apps on my phone. You probably have more than me. My browser bookmarks and history are stuffed with sites I infrequently use. We’re inundated with “apps” that require us to learn some bespoke interface just to get something done. That era is coming to an end. Quickly.
Don’t take my word for it. Many are acknowledging that AI agents have made many apps unnecessary. Oh, I still need the data or function from that app. But I never want to “see” it again. Drop the user interface. Go headless. Why am I still logging into a system three times a year to mark my vacation time? Or navigating multiple travel sites to find the best deals? Or even doing “regular” web searches to find stats to include a presentation? I just want my agent, harness, or super-app (like Gemini, Grok Bot, Meta Muse, Claude Cowork) to do it.
So if we’re not building a bunch of static frontends, what should developers do instead? Here are a few suggestions. Let’s look at two categories of tools that developers should create. And then three types of activities developers (or agents) can do with those tools.
Create CLIs, APIs, skills, and MCP servers for anywhere access to data and functionality
Good models have “computer use” where they can navigate your web frontend if they need to. But spelunking your DOM wastes tokens and time. At least create some WebMCP tools on your site to give agents a better shot at completing a task.
But honestly, those harnesses, super-apps, and agents just want your data and commands. Give it to them! Salesforce announced this headless experience in a partnership with Anthropic (ClaudeForce). Box too. HubSpot as well. Expect most SaaS products to do this. Same with platform companies. Google shipped an Android CLI for use in any harness, 150 agent skills, and tons of remote managed MCP servers. UI optional.
We did before, but now we have even more options to build these types of tools. There are toolkits for creating MCP servers, generating skills, and even cranking out a CLI. Help out your users by giving them the tools they need to get to your data and functionality without forcing them to use your UI.
Create A2UI or MCP Apps components for dynamic rendering
The technology is now there to create more personalized and interactive experiences for humans.
A2UI is a protocol that enables agents to create dynamic user interfaces. You have the data and you have the available widgets that an agent sends a client to render. You’re not shipping arbitrary HTML and JavaScript; the agent sends component descriptions that get safely rendered client side. There are renderers for Angular, React, Flutter, and more. Developers should consider building components that agents can use to compose a beautiful, personalized site for the user.
Then you have MCP Apps. Have you looked at these? Here you’re returning interactive HTML interfaces that get rendered in your agentic chat experience. Developers can build these so that chat users get more than boring text results in their sessions.
Build these portable UI components that users can apply to get richer experiences on whatever surface they can render them on.
Retrieve info or trigger action from where I’m at
Let’s now look at the user.
Maybe I don’t want to use your fancy portal to get my work done. Give me those CLIs, agents, MCPs, or MCP Apps to retrieve the data I want, in the surfaces I’m already using.
Here in my Google Antigravity CLI, I’ve got a reference to the remote MCP Server for Cloud Run. I could also use a locally installed CLI to do the same things.
Using an existing MCP or CLI, I can do most anything with Google Cloud Run. Want a list of running services? Don’t break flow and bounce to a fixed UI somewhere else. Just get the data in your harness or super-app of choice.
Using those tools created by app or product owners, we can now aggregate or interact with systems however we want. Maybe instead of a web front end, you build an agent. Cool. I took a hotel concierge agent that might have warranted a whole fancy website, and used it from within Gemini Enterprise instead.
Need something more interactive? I built an MCP App that talked to Google Cloud Run and returned a fancy dashboard that I can use from within Gemini Enterprise. Why leave my super-app if I don’t need to? First I configured my MCP App (running in Cloud Run) as a data source.
And now anyone can chat with it, getting a full UI back in Gemini Enterprise chat experience.
Build on-the-fly visualizers, apps, and pages personalized to me
The era of personal software is here. Maybe that software is disposable, or just for you. Possibly for a small circle of friends. Whatever. Build entirely new interfaces that work exactly for you by mashing up the APIs, CLIs, and MCPs available to you.
For example, I could use a series of pre-built A2UI components to create a hotel website that reacts to the situation at hand. Completely new user? Show one thing. Checked-in hotel guest, here’s another. Frequent guest coming to book a room? Here you go. Instead of building dozens of static pages to anticipate each scenario, I have ONE PAGE that applies relevant components on the fly.
Build wherever-I-want-it experiences to get work done
This is a variation of the previous two. You might want to stay in your tool of choice, but also want a customized interface just for you. Let’s try that.
We released a new MCP server for the Google Cloud CLI. It executes gcloud CLI commands remotely. This is handy if you’re on an agentic client that can’t execute CLIs, trusted or otherwise.
I could use this to build a full Google Cloud management experience as a Chrome plugin! I can’t easily execute CLI commands in the Chrome sandbox without some hoops. No problem. I can use the logged in user to invoke the MCP server via agent inside a custom plugin. This makes it possible to do almost anything in Google Cloud. from anywhere.
In this example, I spun up a Google Cloud Pub/Sub topic while browsing my daily reading list.
Look, static frontends won’t become truly optional for a while. But developers, don’t wait too long to adapt. These personal agents are growing fast, and fairly soon, many websites and mobile apps will have significantly more agent visitors than human ones. Build accordingly!
Many of today’s items encourage us to do more thinking. Ask more questions. Consider your architectures before unleashing agents. Start with first principles. Hard to disagree with any of that.
[blog] You should all be asking way more questions. This statement applies to virtually every aspect of life. Unlock your inner four-year-old. Ask questions, express interest, pay attention to answers.
[article] Should workers be paid for AI skills? Interesting problem. Might you start paying new hires more for the AI skills you need, without compensating current staff that upskilled to the same level?
[blog] Pub/Sub AI Bytes: Part 3 — Generate and edit SMTs with Gemini in seconds. One of the examples here is notable. Many times, AI isn’t the point. The point is creating something that might have required specialized knowledge before. This shows that AI is helping users create message transformations faster.
I had a relaxing birthday weekend and was raring to go today. Not surprisingly, this reading list has a handful of AI-related items.
[blog] Experts Lead Experts. Obvious? You’d think so. But not everyone believes that experts should lead experts, and we put “people-only managers” in place.
[blog] Do Angular apps need WebMCP? Can’t an AI agent just figure out a web page through the DOM versus needing the tools provided by WebMCP? Sure, but as this test shows, it consumes more time and tokens.
[article] Apps, Agents, and Aggregation. This is all feeding my confirmation bias, but it’s hard not to notice the trend. We’ll use super-apps that aggregate through agents, generative/personalized UIs for the rest.
I wrapped up today a bit sooner to sneak out for a weekend getaway. Inbox is at zero, sun is shining, and next week’s problems will be next week’s problems.
[blog] Stop watching your agent work. I have the opposite problem. I forget that I asked the agent to do something and then I need to catch up on what happened.
[blog] An Ode to Showing Up. It’s not about intensity, it’s about consistency. Show up, and do the work, regardless of how you’re feeling.
When chatting with someone yesterday, I realized one reason this newsletter exercise helps me so much. We all notice a ton of information each day, but don’t always take a moment to imprint any of it on our longer-term memory. Writing this out by hand each afternoon (no AI, I promise) triggers a tangible moment for these various ideas to durably persist in my brain. Even if you don’t do something like this newsletter, take a daily moment to stop and reflect on what you’re learning so that it sticks!
[blog] The Pragmatic Guide to Agent Skills. A year ago, there were no agent skills. Now we probably have too many. But there are credible places to apply them, and Daniela offers some useful instructions.
[article] What makes a great co-worker in an AI-native workplace? What do we want from the AI collaborator in our teams? Do we expect the same things that we expect of human teammates? Looks like the priorities are different.
[blog] Do Frameworks Matter Anymore? Feisty question. You could make a viable argument that we don’t need the structure and bootstrapping help of a web framework. But this post says there’s durable value, even with AI coding at our fingertips.
It’s been one of those days that feels productive, but there have been so few pauses that I need to replay the highlights during the drive home to be sure.
[blog] Launching MCP Apps in Toolbox. Return interactive UI components along with your data? MCP Apps are an interesting idea, and reasonable if your client can render a web view.
[blog] The Internet Is Not Ready for the Agentic Wave. Yes, the web assumes you’re human. Page layouts, CAPTCHAs, carefully crafted calls-to-actions on videos? Agents don’t care about any of that.
[blog] My comment on MCP was always a bad idea? Makes sense. MCP isn’t necessary if your client has full access to CLIs and APIs. But many clients don’t, and won’t.
It was a pretty insane day for frontier models. OpenAI shipped GPT-6 Sol and Luna at very competitive pricing, while Anthropic dropped Claude Opus 5.5 with similar performance to Fable but 40% cheaper. Model-as-a-service stays very interesting when it’s priced competitively to run-it-yourself models.
[blog] Analyzing Jev, a new AI model. Great down-to-earth description of what this model does, and how it probably works.
[blog] Bring Jev to BigQuery with Cloud Run. Extensible systems, for the win. Jeff uses BigQuery’s remote functions capability to invoke the Jev model for fast scoring.
[blog] Trying the Software factory pattern. I liked the callouts to the moving bottleneck, and some of the things that must be true for this pattern to work.
[article] Meta’s Muse is outpacing ChatGPT’s early mobile launch. They’re doing everything right at the moment. Great messaging, solid experience, wide appeal. I’m sure there will be snafus (permissions, etc), but the personal agent moment is now.
[article] Instagram Writing and the Decline of Reading. I’m guilty of this, or maybe as Steve says, just adapting to the times. That said, embrace nuance and long-form reading, not just punchy definitive statements.
[blog] Your Adaptive UI Has No Screenshot. Test the inputs, not the outputs. As we have more generative interfaces that don’t look the same for each person, you’ll want to learn more about how the user got there.
[blog] Colab is now part of your Google AI plan. There’s a heck of a lot of value in these plans. Now you get premium access to better hardware for data science work.
[article] Fixing agent memory. We’re probably hold agents to a higher standard than we hold ourselves. I forget or misremember things all the time. Matt says we need a way continuously validate agent memories.
