Author: Richard Seroter

I ended up with some open time today to build, which felt good. Maybe some more this weekend, but the weather might be too nice to stay inside. Good thing I now have Antigravity Remote Control 🙂

[blog] AI Conversation Trajectories: An Indispensable Primitive & A Terrible Name. We’ve settled on “trajectory” as the name for a completed AI session. Ali hates the name, but recognizes how important this resource is.

[blog] Antigravity Anywhere with Remote Control. So awesome. Now I can access my machines running Antigravity from anywhere. It’s got full awareness of my projects and local context, and I can do most anything from the web interface.

[article] This post will save you tokens. I would behave differently if I was paying for all my tokens right now. Including following some of this advice.

[blog] Building Agents in Go: Where Genkit Shines vs ADK 2.0. Both of these frameworks are great, and heck, both by Google. But there’s a place for each.

[blog] Why I Tell My AI Coding Agent: “Prefer Dart Over Python”. Hear this person out. The argument is pretty good!

[blog] Good writing is obvious, not original. Just write clearly; that’ll stand out more than trying to for originality at every moment. And share things that are obvious to you, but maybe not others.

[blog] Inside the Gemmaverse: Celebrating one billion Gemma downloads. I like that this post celebrates some rad things people built with Gemma so far.

[article] AI productivity gains aren’t translating to IT workload relief. The four-day workweek isn’t on the horizon. As you can get more done, more work shows up.

[blog] Stripe to Acquire OpenRouter: Why Everyone Is Obsessed With Model Routing. A bit of a look at this recent acquisition, but also a nugget about why good routers have data that helps them proactively redirect your requests.

[blog] Bringing Primary Constructors to Dart. Yes, new language features matter, even in the AI-writes-all-the-code era. Easier readability is still important.

[blog] Inside Android Skills – Built for deprecation. I’ve got strong opinions loosely held regarding skills. Useful, yes. But a new type of durable documentation? I don’t think so. I’m more aligned with this post’s POV.

[article] Salesforce wants to move AI coding into a shared workspace with Slack Code. Smart. It’s a place (Slack) where a lot of devs already spend time. Don’t make them switch surfaces. More here, from Slack.

Want to get this update sent to you every day? Subscribe to my RSS feed or subscribe via email below:

I’m adding things to my “you should try this technology” list much faster than I’m getting through it. At some point I’m going to declare bankruptcy and start over. Tomorrow’s looking like a clean “no meeting day” so I stand a chance to get through two of my top ones. Stay tuned.

[blog] Expanding Google Antigravity for enterprise customers. Now Gemini Enterprise users get the Antigravity harness across their favorite IDEs, pooled usage among teams, and backend enterprise controls. More here.

[blog] Practical Loop Engineering. Educational post on the type of agent loops, and how Addy applies them in his own work.

[blog] Add telemetry and observability to your AI agents. Does it really make a big difference to give your agent a memory? Jason built a reference system to test it out and see how behaviors diverge over time.

[article] Your coding agent got the onboarding your developers never did. I chuckled at the headline, because it hit home for me. Are we giving better budget, sponsorship, and patience to agents than we did to humans? Probably.

[blog] How Google Cloud MCPs Give Claude Code Inside Access. Investments in skills and MCPs shouldn’t be tool specific. The whole point is to make sure anyone can use them. Here’s one example.

[article] Only daily AI users feel positively about job security. I guess that doesn’t feel surprising. Those embracing and using the latest thing are likely more future-proofed.

[blog] 20× the CI traffic without getting slower: How we rebuilt Git serving at Datadog. You probably don’t have a git setup as complicated as this. But there are good lessons on unnecessary cloning, recognizing that CPU is the bottleneck, and being smart on your fetching strategy.

[article] Is Your Organizational Culture Too Nice? You know what this looks like. Bland feedback. Everyone gets great performance reviews. Everyone gets a voice. Stop that. Care deeply about others, while aiming for candid, results-focused kindness instead.

[blog] How AlloyDB ScaNN scales vector search to 10 billion vectors. Tremendous performance for a relational database. Here’s how we did it.

[article] When Your Buyer Is an AI Agent. Time to challenge your existing assumptions on pricing and managing buyer relationships.

Want to get this update sent to you every day? Subscribe to my RSS feed or subscribe via email below:

I had jury duty today. Our pool of potential jurors wasn’t needed, so another year without serving. That might be best for society.

[article] Enterprises with AI context layers report agent failures at more than twice the rate of those without one. “Response correctness” is super important to enterprise buyers. And probably all of us. Building these context layers takes time.

[blog] The Unreasonable Effectiveness of Docs in Product Development. Write the docs for the product you want to have, and then start building? There’s something there.

[blog] Harnesses are Situated Agents. The harness is where it’s at, as there’s a lot to manage and those layers are critical.

[blog] Gemini for Go Developers – Part 2: Coding with Gemini. This post will get you all the information you need to use AI to build Go apps.

[blog] Cursor Origin review: An engineer’s perspective. Ready to bounce from GitHub? It’s likely too early to make a big move to Origin, but it looks promising. More about Origin here.

[article] What the Microservices Era Can Teach Us About AI. You might share some lessons, but don’t accidentally treat these paradigms the same.

[blog] Go 1.27 is released. Highly anticipated update! We introduced some great language changes, and improved the already-awesome standard library.

Want to get this update sent to you every day? Subscribe to my RSS feed or subscribe via email below:

Long day. Not done yet as I host my monthly all-hands calls in both AM and PM times to catch our global team. Fortunately these calls are fun, and it’ll distract me from a day of navigating corporate mazes.

[blog] AI usage patterns in software teams. Quite interesting data about the breadth of roles (and levels) using AI regularly. And what’s NOT changed.

[blog] AI Code Review Best Practices. Every team will do their own thing, but here are some viable techniques that work for this team.

[article] Nobody’s Actually Prioritising ‘Value’. Oooh, I like this take. “Value” is contextual to the recipient, and can only be measured after the value exchange occurs. We’re prioritizing work based on confidence in our hypotheses.

[blog] Build zero-trust AI agents with Google’s Agent Development Kit. If you’re using AI to generate AI agents, your tool might skip over this sort of guidance. Don’t make that mistake. Explicitly feed this in.

[article] A better approach to generative UI. Don’t just render out HTML/JS from your LLM. There’s all kinds of risk with that approach. Without saying so, this article makes the case for A2UI.

[article] A Home for Personal Context. Where does context live? Locally on a laptop? On the web? Or in your pocket? Probably all, in some way.

[blog] Google ADK 2 Graph Workflows: A Complete Guide with Code Examples. Romin goes deep here in a very readable post. Define deterministic steps, but still implement a variety of patterns.

[blog] When the Hard Part Stops Being Hard. AI changes how long things take to do. We’re coming to grips with what that means to publication volume, research value, and our intellectual ambition.

[blog] How I Planned an Italy Trip with an Agentic Workflow. As the AI interfaces get simpler, these types of use cases will explode.

[blog] Stop burning tokens on code review. It’s where you’re using the most tokens today in the SDLC. Use a cheaper (local?) model, create custom linters, or design this process with cost in mind.

[blog] Staying Ahead of Adversarial AI Through Agentic Source Code Review. It’s almost guaranteed that you’re piling up more code now that it’s easy to produce it. How are you scanning it at scale for vulnerabilities?

Want to get this update sent to you every day? Subscribe to my RSS feed or subscribe via email below:

It’s “strong opinions Monday” here at the Reading List. Check out some strong takes below.

[blog] A2A joins AAIF’s open agentic stack. We’ve been working on this for a bit. I’m glad to see the open agent-to-agent protocol migrate to this foundation.

[article] When AI Writes the Code, Specifications Need an Exit Strategy. Super interesting post. We haven’t reached industry-wide agreement on the role of specs in the software process.

[article] Research: The Innovation Problems AI Can’t Solve. One thing I took away from this is “don’t rush” and avoid cognitive surrender. Use the AI as a tool, not a substitute for thinking, seeing things for yourself, or taking ownership.

[blog] Antigravity: the busy PM’s best friend. And here are great problems that AI *can* solve. Especially with human spot-checking.

[blog] How Kenn is doing Agentic Engineering. Wonderful post about how this team works with AI. I liked the pushback on loops and dark software factories. Those things are working for some people, but it’s overhyped.

[blog] Skills Sprawl: When Too Much of a Good Thing Confuses Your AI Agent. You’ve gorged yourself on skills. Step one was recognizing it. Step two is to do something about it. Excellent post from Darren.

[blog] The SKILL.md Fallacy: Phase Transitions & Process Isolation in Coding Agents. What a write-up by Ali. He makes a strong case that we’re misusing skills and should be looking at subagents instead.

[blog] Fabled Too Hard? How I Stay Engaged in AI Development. Don’t lose your creativity or your agency just because you do AI-assisted development.

[article] Elevating Antigravity agent skills, Part 5: Subagent management. maybe you follow Ali’s above advice, but get good at learning how to manage your subagent teams.

[article] GitHub outage disrupts developers worldwide in latest setback for Microsoft coding platform. Oof. Rough day for GitHub on an issue that’s plagued them for a while now. Timing wasn’t great, as a legit competitor launched today.

[blog] The Shapes of Agent Memory – Files, Stores, and Experience. I’m not smart enough to extract all the learnings from this, but sounds like databases are better that text files when AI needs to process many conversations.

Want to get this update sent to you every day? Subscribe to my RSS feed or subscribe via email below:

There’s a lot of agent content today. But it’s about fairly leading-edge topics, so give them a scan.

[blog] Introducing Neocloud And NeoPaaS: The Next Frontiers Of The AI-Native Cloud. Do we need new terms? Our analyst friends think so. And this doesn’t even include the “accidental clouds” like xAI and Meta.

[article] The Two Pillars of Post-training: Reinforcement Learning and Supervised Fine-Tuning. Here’s another great post-training article in this series. If you want to understand how models get better without training an entirely new model from scratch.

[blog] Speeding up generative UI with async A2UI. This is top-2 on my “need to play with” list of technologies. Andrew is ensuring it stays there.

[blog] Patterns and problems in emerging multiagent systems. What a fascinating read from Anthropic. Looks like agents don’t naturally coordinate any better than humans do.

[article] Can software factories actually work? You’ve heard the term, maybe you’re even building (or running?) one of these. But are we capable today of running lights-off software lifecycles where AI does literally everything from deciding what to work on, to shipping the final software?

[article] How does AI-assisted coding impact code comprehension? We comprehend less when the AI does more. Not shocking. There are ways to mitigate, as called out here.

[blog] Building Deterministic Agent Graphs with ADK 2.0. We don’t have to leave every decision up to the agent. Steering, along with persistence, is helping agents grow up.

[blog] Agentic Engineering Is Just Everything We Haven’t Been Doing. Are we rediscovering (or just getting reminded of) work that is a traditional part of a given workflow?

[blog] AGENTS.md vs. skills: How to steer a coding agent. Each operate at a different level, and steers in different ways. Eval this yourself.

Want to get this update sent to you every day? Subscribe to my RSS feed or subscribe via email below:

Google is a big, successful company, so I can’t really play the “underdog” card. But I’m glad to see us rocket back to (influencer) relevance with today’s Gemini 3.7 Flash launch. Great model.

[blog] Introducing Gemini 3.7 Flash. Surprise! We just shipped Gemini 3.6 Flash, but we’re picking up the pace. And this one is notably better at coding and agentic sessions.

[blog] Reviewing code is a skill. Definitely. It’s something you can get better at my practicing, and something that can be taught to others.

[article] Netflix Adopts Cloud-Native Job Queueing System Kueue to Replace an In-House Solution. This subsystem has been around nearly four years, and I’m happy to see major adoption of it.

[article] Introduction to Post-training. Very educational post. I’m far from an expert (or even functional) on the details of pre and post training.

[blog] Vertical slices, their ownership and external dependencies. I like this pattern for building, but there are ways it goes sideways. This post looks at how to manage externa dependencies.

[blog] Loop driven development with Antigravity CLI. Shocking to say, but sometimes you should just use code instead of twisting your AI harness to fit every task.

[article] Google says Go is well suited to AI-generated code. That’s true, we do say that. I got a lot of value from reading the various comments (positive and negative) to our push on this topic.

[article] How Today’s CEOs Can Prepare for Tomorrow’s Crises. I. like the point that we’ll never be an expert at everything, but we can be good at asking the right questions.

[blog] WIBTM: Give an agent access to memories created with AI dreams. Session telemetry (and trajectories) become “dreams” that fold into long-term memory. Comes in handy for later sessions.

[blog] Serverless LLM Blueprint: Build a Fine-Tuning & Serving Platform on Google Cloud. Do serving yourself—as in, not a single managed API—but with entirely managed services that only charge compute when in use? Good pattern.

[blog] The Artifact is Free, Assurance is the Product. Indeed, but even then, the provider can only assure so much.

[blog] Bring your spreadsheet data to life with Sheets canvas. Why build an app or dashboard separate from your source data (spreadsheet) if you can just do it all together? Pretty sweet experience.

[article] Technical Debt in the AI Era. Good research piece that asks whether our prior approach to technical debt still matters now that we have AI. Short answer: yes.

Want to get this update sent to you every day? Subscribe to my RSS feed or subscribe via email below:

Day baseball games for my Padres are a problem because I keep an eye on the score (sometimes via a muted television). It confuses people if I cheer randomly during business meetings.

[article] Some Simple Economics of Open versus Closed AI. Thoughtful piece, with relevant historical references. Idea compounding and complementary assets will likely dominant AI in the years ahead, not monopolies on ideas.

[blog] No, local models will not win. A high-quality hot take here. Local/open models aren’t necessarily cheaper and rarely more efficient. They still have a clear value for a wide variety of use cases. But it’s probably naive to say that the future of AI is open source. It’s more complex than that.

[blog] I built the same agent three ways: Interactions API, ADK, and the Antigravity SDK. I’ve been looking at the same things lately (plus Genkit). It’s not about identical ways to do the same thing, but choosing software attributes and the right approach to achieving them.

[blog] Introducing the Developer Device Platform for agentic mobile app development. Get access to a Android devices in the cloud so that you can test your mobile apps? Neat. I can’t even access this myself quite yet, but I assume we’re rolling it out in stages.

[blog] That Gap Is Still Widening, But The Bottleneck Was Never Engineering. I don’t think this is a controversial take. Giving your engineers better tools doesn’t change the work around the work that slows down delivery.

[blog] What’s new in Flutter 3.47. This outstanding multi-platform framework gets great updates for developers. Also, a corresponding Dart 3.13 release.

[article] TDD inside the agent loop – theater or actual value? Bringing every “classic” process to new tools and systems? Maybe you’re fighting an (irrelevant) uphill battle. This was a good test.

[blog] The Appwrite CLI is now written in Go. Their CLI is now faster, smaller, less resource intensive, and simpler.

[blog] Mastering Edge AI on Raspberry Pi with LiteRT and Gemma. Isn’t it amazing what you can do nowadays with ultra-lean footprints?

[blog] Where an AI Watermark Can Hide in Plain Text. How in the world do you watermark text? Anthropic is doing that with Claude now. Google DeepMind showed an example of this years ago.

[blog] Introducing Custom Agents. This is. a great new feature in Google Antigravity. Define custom agents and use as the primary or sub agent in your builder workflows.

[article] InfoQ Cloud and DevOps Trends Report – 2026. It’s weighted towards early and late majority items. But those left-side, innovation-focused items are fairly critical for those working with modern workloads.

Want to get this update sent to you every day? Subscribe to my RSS feed or subscribe via email below:

I spent time with a great company who is trying to bring AI into more of their product management function. Not easy, but high impact if you get it right.

[article] Can Agents Use a Computer Yet? We’ve Got the Data. You won’t find clean APIs for every system. Sometimes you need the AI to act like a user to “see” what’s going on in the software. That capability keeps getting better.

[blog] A Fresh Definition of The Product Role. Just because everyone can build, doesn’t mean everyone can build the right thing. Product thinking is a real thing that people work to develop over time.

[blog] Why Go is an Ideal Language for AI-Assisted Software Engineering. Cameron and I make the case. And it’s a heckuva strong case.

[blog] Expanding connectivity in the Americas: Introducing Alisios, Canoa, and OlaLuz subsea cables. More connectivity means better opportunities for more people.

[blog] Reflections on Ai4 conference – some absences from a tech writer’s POV. Good insights from Tom. The “missing” topics at this event are surprising.

[article] Agentic Code Quality. This piece from Addy challenged me. His argument is strong. I think I struggled with the idea that quality could be completely codified into agent constraints. I guess there’s no reason that’s not possible.

[blog] Gemini for Go Developers – Part 1: The Gemini Model Family. Great introduction for people who are starting to figure out how to bring AI into their apps.

[blog] How I use LLMs to learn complex topics. Great little example. Learn by asking models to build out some foundational knowledge on a topic, ask it to review for accuracy, and then build a simulation of that topic.

[blog] Expanding the Google AI Professional Certificate with vibe coding. This wildly popular cert has fresh curriculum for builders. Give it a try.

[article] Are agents really killing UI? It’s all headless APIs now? Nope. We’ll have UIs around for a long time, with generative UIs and personal software on the upswing. But, anything your human users can do, an agent should be able to do too.

Want to get this update sent to you every day? Subscribe to my RSS feed or subscribe via email below:

Great weekend, and great Monday. Today’s reading list has some bangers on it with some hard truths to digest.

[blog] Introducing Muse Glimmer: An Open Agentic Model That Runs on Your Device. Some great model drops by Meta lately. This one is small and agent-ready. More here.

[blog] “Code was never the hard part” is an insult to all programmers. Agree. Stop saying this. Coding isn’t easy. If “figuring out what to build” is the hard stuff, how come PMs aren’t the tech superstars? Some honest perspective here.

[blog] PDFs are terrible. Amusing anecdote where Accenture realizes that non-engineers are eating a lot of tokens doing things like converting PDFs to easier file formats.

[article] Managing 8 engineers now feels like managing 4 teams. It’s a new challenge for managers, but the point here is that product management just became the new bottleneck.

[blog] Decoupling in Software Architecture Moves Complexity. Right. You’re not picking up simplicity as you distribute more components of your system. Other benefits, sure, but the complexity remains.

[article] What are code reviews even for? Yes, dramatically increasing the volume of code while retaining the same code review process is unsustainable. Something has to change. Just make sure you don’t lose an important means of knowledge transfer or building ownership.

[blog] Agents Don’t Magically Understand Your API. You still need good API design, even if AI is the API consumer.

[article] Why observability doesn’t explain what happened. The piece argues that observability tools tell you what’s going on in your system right now. But “why is this happening” and “what triggered it” are likely sitting outside its reach.

[article] Elevating Antigravity agent skills, Part 4: Subagent messaging. This type of session messaging is new for Claude, but it’s been part of Antigravity for a while. James has a good piece with advice and anti-patterns.

[article] Platform Engineering ROI: What it costs to build your own platform. It’s usually more than you think. Sometimes less, but it’s easy to leave out hidden costs that pile up or distract your team with undifferentiated work.

[article] Why Open Source Matters for AI. It’s an “and.” I don’t mind closed models, and many users don’t either. But open weight models make a whole additional set of use cases possible.

[blog] Open-source is NOT the same as open-weight. Virtually no one gives you an open source model. You can’t change the raw data or pre-processing. All you get is frozen weights to manipulate.

Want to get this update sent to you every day? Subscribe to my RSS feed or subscribe via email below: