🎧 Listen to this episode

📰 Original post by Richard Seroter

Nine articles today, covering agent protocols, production observability, vibe coding, front-end architecture, and some good old-fashioned engineering wisdom.

Richard kicks off this edition with a reflection on how fast the thinking around AI agent skills is evolving. Anthropic, OpenAI, and Google all published guidance this week on how to write effective skills for coding agents, and Richard’s take is refreshing: build things and figure out what works for you.

First up, Charity Majors at Honeycomb wrote a piece called “Production Is Where the Rigor Goes.” This one is a response to the recent Deer Valley unconference, where Martin Fowler and Thoughtworks brought together some of the best minds in software to discuss engineering in the AI era. They asked the big question: as AI generates code faster, where does the rigor go? The group identified five places — spec review, test suites, type systems, risk mapping, and continuous comprehension. But Charity noticed a glaring omission: production itself. She argues that production is reality. Observability is just another word for understanding, and the best way to build good software has always been to ship, observe, and iterate. AI makes this even more urgent because as the rate of change goes up 10x or 100x, all the duct tape and tribal knowledge that held systems together simply won’t scale.

Next, Google announced a major upgrade to the vibe coding experience in Google AI Studio. They’re calling the new coding agent “Antigravity,” and it turns prompts into real, production-ready applications. We’re talking multiplayer experiences, Firebase integration for databases and authentication, support for React, Angular, and Next.js, and a secrets manager for API keys. The agent can detect when your app needs a database and provision one automatically. Google says hundreds of thousands of apps have already been built internally with this tool.

From InfoWorld, a thoughtful essay titled “We Mistook Event Handling for Architecture.” The argument is that front-end development became so sophisticated at handling events, dispatching actions, and coordinating reactive pipelines, that we started treating event handling as the architecture itself. The Redux era reinforced this. But the author makes a distinction: events tell you what just happened, while architecture should describe what is true right now. A quieter shift is emerging toward state-first thinking, where dependencies are declared explicitly and derived values are functions of underlying state. Events still matter, but they return to their proper role as inputs, not the backbone of the architecture.

Here’s a fun one from Google Cloud. An engineer used the Gemini CLI to test whether large language models follow Benford’s Law. In many naturally occurring datasets, about 30% of numbers start with the digit 1, and less than 5% start with 9. It’s used in forensic accounting to detect fraud. The experiment asked Gemini to generate populations for fictional cities and checked whether the leading digit distribution matched. The result? Not even close. LLMs have internal biases that produce unrealistic distributions. You can calibrate by putting the expected distribution in the prompt, but the real takeaway is practical: treat LLMs as planners. Let them define the parameters, but use actual code to generate the numbers.

From Harvard Business Review, an article asks what the best AI users do differently. Organizations are investing heavily in AI tools but struggling to measure whether they’re actually improving quality, speed, and judgment. The research finds that the most sophisticated AI users tend to be at the manager level and above. It frames AI adoption as a skill development challenge, not just a tooling problem.

Google’s developer blog published a comprehensive Developer’s Guide to AI Agent Protocols. If you’ve been confused by the alphabet soup of MCP, A2A, UCP, and others, this is the piece that makes sense of it all. They walk through building a restaurant supply chain agent step by step using Google’s Agent Development Kit, starting with a bare model that hallucinates everything, then layering in protocols one by one. MCP handles data access, A2A enables agent-to-agent communication via published agent cards, UCP standardizes transactions, and they also cover AP2 for authorization and AG-UI for streaming frontends.

Sticking with agent skills, the OpenAI Agents SDK team published how they use Codex to maintain their own repositories. They keep repo-local skills in a .agents/skills/ folder, with an AGENTS.md file that tells Codex when to invoke each skill. Skills cover code verification, doc syncing, release preparation, and PR drafting. The results: 457 pull requests merged in three months, up from 316 in the prior period. The key insight is that skills work best as narrow, repeatable workflows with clear triggers and concrete outputs.

From the Manager Dev newsletter, seven unwritten laws of software engineering that every developer learns the hard way. When production breaks after your deploy, just roll back first, then debug. Backups aren’t real until you’ve actually restored from them. Logging is perpetually imperfect. Every external dependency will fail. And the best one: there is nothing more lasting than a temporary fix.

Finally, Google shared updates to the Universal Commerce Protocol, their open standard for agentic shopping. New capabilities include a Cart feature for multi-item adds, a Catalog capability for real-time product details, and Identity Linking for loyalty benefits. Partners like Salesforce, Stripe, and Commerce Inc are implementing UCP.

If I had to pick a theme, it would be the tension between automation and understanding. AI is making it faster to build, deploy, and transact, but the articles that stand out most today are the ones reminding us that speed without comprehension is just organized chaos.

Articles Covered

1. Production Is Where the Rigor Goes — Honeycomb
2. Introducing the new full-stack vibe coding experience in Google AI Studio — Google
3. We mistook event handling for architecture — InfoWorld
4. Do Large Language Models follow Benford’s Law? — Google Cloud / Medium
5. What the Best AI Users Do Differently — Harvard Business Review
6. Developer’s Guide to AI Agent Protocols — Google Developers
7. Using skills to accelerate OSS maintenance — OpenAI
8. The unwritten laws of software engineering — Manager Dev
9. AI shopping gets simpler with Universal Commerce Protocol updates — Google