AI's Hidden Costs: Debt Accumulation and Coding Pitfalls
I like AI and use it, but we can also do worse work and terrible things with it. Two articles I read today: What kinds of new debt are teams accumulating with AI? https://t.co/9SRgF4KdgU Some uncomfortable truths about AI coding agents https://t.co/pfPI2yTMEO
Legacy System Migrates to Google Cloud, Serving 160M Users Faster
There's a decent chance you have an https://t.co/A6q3Fkidky account for digital interactions with the IRS, DMV, and popular ecommerce sites. They moved from legacy architecture to @googlecloud and supports 160 million members with a better/faster stack. https://t.co/pFGhUtVgaQ https://t.co/hljWAZ1rnO
Secure Hybrid Self‑Managed and Managed MCP Server Setup
You could use a mix of self-managed and managed MCP servers. Here's an example of using both, and securing them in a production-ready way. https://t.co/reHeaq6QEV https://t.co/5pLxHwGKWv
AI Splits Firms Into Fast Vs
"Instead AI is splitting enterprises into fast-learning and slow-learning teams and is rewarding organizations that redesign work, govern risk, and turn lower software costs into more software, not less." - @mjasay https://t.co/DpencJcAo1
Stateless Semantic Search in BigQuery Simplifies Small Datasets
You have any giant, convoluted code or SQL logic to handle data values that might be similar? @JeffONelson did. But he shows off a new stateless semantic search in @googlecloud BigQuery that might be a lifesaver for small datasets. https://t.co/RU5q8SoJb4
Use AI as Tutor, Not Shortcut, With Guardrails
"How much should I use AI? What should I use it for?" This research project with students at UC Berkeley revealed some great insights: Use as a tutor (not a shortcut), build guardrails to protect learning, know when to turn...
Open LLMs Power Medical Apps and Self‑Hosted SRE Agents
You can do all sorts of things with open LLMs. We just announced the winners of a competition to build prototype apps using our open medical model, MedGemma: https://t.co/ivRNKsJloa You can also use Gemma to build your own self-hosted SRE agent: https://t.co/bssaP1Sfh3
Firebase Delivers Full Backend for Flutter Apps
If you're building mobile or web apps, setting up the backend—auth, databases, file storage, analytics, push notifications, etc—is legit work. This post points out why having a @Firebase backend for @FlutterDev apps gives it all to you out of the box. https://t.co/BOOaxRnb9g...
Visual Guide Demystifies CPUs, GPUs, TPUs, NPUs, LPUs
CPU vs GPU vs TPU vs NPU vs LPU https://t.co/SdMvlqSkSK < that's a lot of three letter acronyms. The visual explanations in here helped me. https://t.co/JMCeP5384p
New Deployment Adapter API Lets Next.js Run Beyond Vercel
Next.js is popular, but not easy to deploy outside of Vercel. But thanks to teamwork across companies, there's a new stable Deployment Adapter API. And @Firebase is deeply involved and giving you a great way to run these apps. https://t.co/8Nk7k90yTY
Great DevEx Turns Friction Into Seamless Flow
From Friction to Flow: How Great DevEx Makes Everything Awesome https://t.co/mInTjuCShW < I'm so glad I get to work with @nicolefv and learn from her. Everyone gets her wisdom in this @InfoQ video/transcript.
Java Powers AI-Generated Lyrics and Music with Lyria 3
Make music with ... Java? @glaforge wrote up a great example of how to use the new @GoogleAI Lyria 3 to create lyrics and music from a Java application. https://t.co/OIKEDjtyu0 https://t.co/zFBUKSi8hE
Scion: Preston's Self‑Organizing Agent Orchestrator, Local & Remote
Preston is the brains behind Scion, this self-organizing agent orchestration tool. Run local, remote, or both. Give us feedback if you try it out. I'm aiming to take it for a swing this weekend.
New Gemini Skill Dramatically Boosts Model Performance
The difference between a "vanilla" request to the Gemini model and enabling this new skill? Pretty dramatic. More to do, but we'll all keep learning the best way to apply these to our agents and tools. https://t.co/Qh459zzQQr https://t.co/MUppwHGUry
Scion: Open‑Source Multi‑Agent Orchestration for AI Swarms
Am I supposed to talk about this yet? It's Friday, let see what happens. We quietly open sourced Scion, a new multi-agent orchestration tool for deploying and managing swarms of containerized AI agents Describe the rules, agents self-organize. All in...
Cut Monorepo Size, Slash Clone Times and Timeouts
By shrinking their monorepo from 87GB to 20GB, clone times dropped from an hour to under 15 minutes, onboarding became faster, CI pipeline started faster, and they saw fewer timeouts. The story from @Dropbox engineering ... https://t.co/XLOcCvbJeu
Emotional Fulfillment Beats Pay in Talent Retention
"Emotional needs, such as feeling valued and supported, mattered more than functional needs like pay, benefits, and hours in retaining and supporting talent." https://t.co/xLzSKDM6Fz < we all crave joy and appreciation in our work
Vibe Coding XR Frees XR Creation for All
"Vibe Coding XR marks a pivotal step toward a future where spatial computing is limited not by technical expertise, but by creativity." https://t.co/qHpu9QOqfC < what happens when any of us can vibe-code some extended reality apps? Cool @GoogleResearch work. https://t.co/LWJ3LtvHmq
Choose the Right Branching Strategy: Trunk vs Feature
Feature branches? Trunk-based development? How do you think about your branching strategy for source code? Here's a good look at some proven patterns: https://t.co/8o8FCGewx8 https://t.co/QYIYdtQ1eX
Five Ironwood TPU Optimization Strategies for ML Engineers
A developer’s guide to training with Ironwood TPUs https://t.co/eFGnybvR0G < it's really five optimization strategies for ML engineers. Check it out.
Pick the Right Deployment Strategy, Even with AI
I know we're somehow now talking about AI agents that just ship code directly to prod whenever, but yeah, you may still care about your software engineering rigor. Here's a post with common deployment strategies and how to choose the right...
2026 Context Engineering: Key Patterns for LLM Knowledge
State of Context Engineering in 2026 https://t.co/MbilPPPrmV < this is good stuff. What are the patterns and considerations for giving your LLM or AI agent what it needs to know? https://t.co/SDHWYHBD6i
Design Agent Harnesses, Manager Timing, and AI-Generated Music
Seroter Daily Reading List – March 25, 2026 (#749): Today’s links look at designing agent harnesses for long-running app development, when a manager should step in with their team, and how to generate legit music with AI. https://t.co/u8kZPntfS8
TurboQuant: Theory‑Based Compression for LLMs and Vector Search
"TurboQuant" sounds like a midtier Marvel superhero. But no, it's the name of a "theoretically grounded quantization algorithms that enable massive compression for large language models and vector search engines" from @GoogleResearch. https://t.co/wCk8aPqOB6
Managers: Know When and How to Provide Direction
If you're a manager, when do you step in and specifically give direction to your team member? And how? For all of us with managers, when do we prefer they engage? I found this article on the topic thought-provoking ... https://t.co/w02M6tOva1 https://t.co/6C3RaNYVQA
Turn Bulky MCP Servers Into Lightweight Binaries for Agents
Maybe flip some heavy MCP servers to tiny, fast binaries that your agent can use within a Skill? That's what @iRomin did here as an experiment and I think the approach has merit. https://t.co/if8SC0djvj
Read Real Books, Not Just Tweets and Listicles
We need to read more. I'm not talking tweets, listacles, or AI-generated articles. Books. Real ones. On a variety of topics. I agree with everything in this piece by @BStulberg ... https://t.co/no90lvjXU6
Threat Handoffs Now Occur in Seconds, Not Hours
"In 2022, the median time between an initial access event and the hand-off to a secondary threat group was more than 8 hours. In 2025, that window collapsed to just 22 seconds." https://t.co/gjePO94A0N < important security data in this new...
Google Adds Dark‑web Intel for Faster Threat Detection
"To get teams the critical data they need to make quick, accurate decisions about rising threats, we’re introducing a new dark web intelligence capability in Google Threat Intelligence." https://t.co/qGKDWJjI36 < identify risks faster and get ahead of adversaries
Google Cloud Vertex AI Delivers Fastest Anthropic Model Performance
If you want the lowest latency and most throughput for @AnthropicAI models like Claude Opus, you should access them from @googlecloud Vertex AI. Don't take my word for it. @OpenRouter data makes the case (h/t @ivnardini): https://t.co/k36mJMakpu https://t.co/jFSGNsptxz
Key Differences Between Agent‑Native and Cloud‑Native Infrastructure
We just snuck out this cool little @googlecloud paper that spells out what to think about when preparing agent-native infrastructure. What changes from a cloud-native approach? Direct link: https://t.co/10AiOuo7LG https://t.co/7Eg6IqHaJa
Modernize Legacy Apps with Stitch and Google AI Studio
Got some functional but tired-looking apps laying around? Sure, we all do. @kweinmeister took one and used @stitchbygoogle along with the revamped @GoogleAIStudio to modernize his architecture and make the app look fantastic, and run on a modern host. https://t.co/QiUcjkY1jB https://t.co/mBtY82xW39
New Short Videos Teach Advanced AI Agent Design Patterns
Many of you prefer learning through short, informative videos. We're doing more of that. I'm enjoying this smart content on agent building from @anniewangtech. AI agent design patterns https://t.co/2mT2PS9hgp Advanced design patterns for dynamic agents https://t.co/IImACq8Dq8
Kubernetes 1.36 Adds Native Scale‑to‑zero Pods
If only my Kubernetes pod could scale to zero. That'd be great for staging/test environments or irregular production workloads. Oh, that's coming in Kubernetes v1.36 after sitting in alpha for years? Sweet. https://t.co/klzNA6Hs0X https://t.co/Yee1mpTLEh
Elastic Read Pools Simplify Scaling in Cloud SQL
I like these new autoscaled read pools for @googlecloud SQL. You might scale your relational database through read replicas, and now you've got an elastic pool that exposes a single read endpoint so your code doesn't have to change. https://t.co/e1YDOUEGRr https://t.co/pq2gEqxcfJ
Universal Commerce Protocol Simplifies AI Shopping Experience
AI shopping gets simpler with Universal Commerce Protocol updates https://t.co/aqwpcfY5zc < new options, capabilities, and identity-linking features
LLMs Show Unexpected Deviation From Benford’s Law
"Benford's Law" says that 1 and 2 are overrepresented as the starting number for most everything. 8 and 9, rarely. @kweinmeister wondered if LLMs follow this statistical law when generating data. https://t.co/hgtjs24FJg https://t.co/hij3SdDHlb
Production Is Real Flight; Observability Guides the Journey
"Formal methods and test suites are flight simulators. Production is flying the actual plane. Observability is how you fly it." -@mipsytipsy Important post from Charity on why you must treat production as more than a place you go to fix...
Built‑in Tools Cheapest; MCP+Skill Most Efficient
Need to control context consumption for your custom-built ADK agent? What do you use? Built-in tools? Skills? MCPs? In my tests, built-in tools alone solved the job at lowest cost. MCPs did well (with more tokens), but MCP plus Skill...
IoT Cuts McDonald's Ice Cream Outage From 19 Clicks
"The ice cream machine is broken." Those are six very depressing words. McDonalds had a 19-click process for marking ice cream as unavailable. Now with IoT and tech solutions, there's failure detection and response. https://t.co/w6mR84wNiX https://t.co/2ufj3Zp3MB
Gemini API Merges Function Calls with Built-In Tools
YES. Instead of separate calls or agents, now you can combine function calling with built-in tools (like Google Search) in a single Gemini API call. This simplifies a few of my own agents. https://t.co/R97to7Af8U
Coding Agents Combine LLMs, Prompts, and Tools in a Loop
A coding agent is basically an LLM + system prompt + tools in a loop. @simonw spells this out in another great guide about how coding agents work. https://t.co/ywlRRAU3VY
Exploring Subagents, Enterprise AI, and Agent Protocols
Seroter Daily Reading List – March 17, 2026 (#743): Today’s links look at what subagents are all about, what the state of AI in the enterprise is, and how to understand all the agent protocols. https://t.co/ELYoylGwu6 https://t.co/49pUY1ToOU
Run AI Inference Across GKE Clusters, Any Region
For those running AI/ML models on Kubernetes, do you feel pinned to one region? Not great. We just previewed the @googlecloud multi-cluster GKE Inference Gateway. Scale inference workloads across clusters, even across regions. https://t.co/t6vL4a7ZEH https://t.co/MTpFrlroKp
Banks' Legacy Costs Stall AI Innovation
Lots of corporate dollars still go to maintaining legacy systems. It's hard to redirect that spend to new AI work. Banks feeling it, and not yet delivering the (AI) products that customers want ... https://t.co/N4HWKA2UML
Colab Becomes Open, Extensible Host for MCP Agents
"By establishing Colab as an open, extensible host, you can now treat Colab as an automated workspace for any MCP-compatible agent." https://t.co/YEXI75WM1d < use the new @GoogleColab MCP server to offload to the Colab runtime, and use notebooks as a...
Prioritize Token Efficiency in Your AI Agents
I see many of you trying to precisely track token usage in your agentic dev tools. But how about the agents you build? It matters more there, where smart token usage has bigger impact. I looked a few scenarios and which was...
Tiny Fraction of Models Capture Half of Downloads
"The ecosystem remains highly concentrated. Approximately half of the models on Hugging Face have less than 200 total downloads, and the top 200 most downloaded models, or 0.01% of models, comprise 49.6% of all downloads." https://t.co/N5AuzECfqO < tons of data...
Enterprises Lag in AI Work Redesign, Prioritize Sovereignty
Kudos to @Deloitte for offering the "State of AI in the Enterprise" report without a reg wall. Findings? Most companies haven't started redesigning work for AI. Sovereignty is playing a big part in vendor selection. Few companies have agent governance. https://t.co/LyqwItP1Hv https://t.co/C7TIt6UGVE
Exploring Gemini API with Nano Banana 2 Guide
Developer Guide: Nano Banana 2 with the Gemini Interactions API https://t.co/4gXFg64W7A < I need to spend more time with this API. @_philschmid has inspired me to invest more in this unified and feature-rich Gemini API. https://t.co/TFxiwComyq