🎧 Listen to the audio summary

📰 Original post on seroter.com

Welcome to Seroter’s Daily Reading, number 751, from March 27th, 2026. Today’s list has eleven links, and Richard kicks things off by mentioning it’s baseball’s Opening Weekend and he’ll be at a San Diego Padres game. Let’s get into the articles.

First up, a really practical piece from the PostHog team titled What We Wish We Knew About Building AI Agents. This one’s worth your time if you’re anywhere near the build-versus-buy decision for AI agents. Their core insight is that before you build a custom agent, you should seriously ask whether an MCP server would do the job. PostHog found that 34 percent of dashboards created by AI were done through their MCP server, which is nearly as much as their built-in agent. They went through three iterations of their agent harness before landing on the Claude Agent SDK with MCP tools and skills. The lesson that stands out: don’t burn innovation points on the harness itself. Use an existing SDK, make MCP the canonical interface, and focus your energy on context — the unique blend of your product’s data and capabilities that no generic tool can match.

Next, a research analysis from the RDEL newsletter asking how Google turned AI IDE features into measurable productivity gains. The big takeaway here is that model quality alone wasn’t the bottleneck. Google studied two AI-powered IDE features across 36,000 developers over 14 months. They found that UX changes rivaled model improvements in impact. Adding a floating button next to code selections increased first-time users by 64 percent. Improving diff rendering cut review time by 7 percent. And here’s a fascinating detail: caching alone boosted the acceptance rate by 17 percent and increased the share of ML-written code by 41 percent. They frame this as an “opportunity funnel” where value leaks out at every stage between prediction and developer action.

Third, a Harvard Business Review piece arguing that leaders underestimate the value of employee joy. Companies have gotten incredibly good at understanding their customers — mapping journeys, using predictive analytics, studying behaviors — but when it comes to understanding their own employees, most still rely on intuition or generic surveys. The disconnect is real: the people who create customer experiences are the ones being least understood.

From the Google Developers Blog, there’s a post about closing the knowledge gap with agent skills. This is about the Gemini API developer skill, a lightweight markdown-based way to teach coding agents about your SDK. The results are striking: Gemini 3 series models jumped from a 6.8 percent baseline to excellent pass rates once the skill was added. But here’s the nuance — older models with weaker reasoning barely benefited. They also acknowledge that Vercel’s research found direct instruction through AGENTS.md can outperform skills, so this space is still evolving.

Gergely Orosz from The Pragmatic Engineer tackles whether the Forward Deployed Engineer role is becoming less desirable. Job postings for FDEs grew more than tenfold in 2025, and company earnings calls mentioning the role jumped from 8 to 50. But only about 10 percent of engineers actually want the job. The reality is that FDE work often ends up looking more like sales engineering or consulting than real software engineering. One developer quit after four weeks, calling it a “typical IT services mindset.”

Then there’s a fun one from Guillaume Laforge about generating music with Google’s Lyria 3 model using the Gemini Interactions Java SDK. Lyria 3 comes in two flavors: a clip model for 30-second pieces and a pro model for full songs up to three minutes. You can control structure, provide your own lyrics, or request instrumental only. The API returns native MP3, so there’s no complex decoding needed.

The Dropbox engineering team shares how they reduced their monorepo from 87 gigabytes down to 20 — a 77 percent reduction — cutting clone time from over an hour to under 15 minutes. The culprit was Git’s delta compression heuristic, which uses the last 16 characters of a file path to pair files for comparison. Dropbox’s internationalization files had the language code earlier in the path, so Git was computing deltas between completely unrelated language files. The fix required working directly with GitHub, since server-side repacking determines what clients actually download.

From the Firebase blog, an update about Next.js deployment adapters. With Next.js 16.2, there’s now a stable Deployment Adapter API — the result of a collaboration between Firebase, Vercel, Netlify, Cloudflare, and AWS. This replaces the fragile reverse-engineering that hosting providers previously had to do to support features like Incremental Static Regeneration.

There’s an article from Telerik about the top three everyday Angular tasks that MCP servers can automate: scaffolding components, turning data into UI layouts, and iterating on existing interfaces. The key point is that without MCP, AI assistants generate code without knowing your Angular version, your component library, or your project structure. With MCP, the assistant connects to your actual tools and generates valid, grounded code.

Wiz introduces their new agents and workflows platform for cloud security. They’ve built three specialized agents: Red Agent for offensive security, Blue Agent for defensive investigation, and Green Agent for resolution. The workflow engine lets teams orchestrate how these agents operate — automatically triggering fixes for high-confidence findings or routing lower-confidence ones through human approval.

Finally, Nicole Forsgren’s InfoQ presentation titled “From Friction to Flow: How Great DevEx Makes Everything Awesome.” Her central argument is that AI has made code generation fast but hasn’t fixed the rest of the software delivery pipeline. She references eye-opening numbers: McKinsey found 40 percent of developer budgets go to avoidable rework, and one estimate puts the cost of technical debt at 1.52 trillion dollars. Her DevEx framework focuses on feedback loops, flow state, and cognitive load.

Articles Covered

1. What We Wish We Knew About Building AI Agents — PostHog Newsletter
2. How Did Google Turn AI IDE Features into Measurable Productivity Gains? — RDEL Newsletter
3. Leaders Underestimate the Value of Employee Joy — Harvard Business Review
4. Closing the Knowledge Gap with Agent Skills — Google Developers Blog
5. Is the FDE Role Becoming Less Desirable? — The Pragmatic Engineer
6. Generating Music with Lyria 3 and the Gemini Interactions Java SDK — Google Cloud on Medium
7. Reducing Our Monorepo Size to Improve Developer Velocity — Dropbox Tech Blog
8. Next.js Deployment Adapters: A Bright Future for Next.js on Google Cloud — Firebase Blog
9. Top 3 Everyday Angular Tasks That MCP Servers Automate for You — Telerik Blog
10. Introducing Wiz Agents & Workflows: Security at the Speed of AI — Wiz Blog
11. From Friction to Flow: How Great DevEx Makes Everything Awesome — InfoQ (Nicole Forsgren)