🎧 Listen to this episode

📰 Source: Daily Reading List — March 23, 2026 (#747) by Richard Seroter

Welcome to Seroter’s Daily Reading, your audio digest of Richard Seroter’s reading list. This is number 747, from March 23rd, 2026. Richard posted this one from San Jose, ahead of an in-person rehearsal for the Google Cloud Next developer keynote. Eleven links today, and a clear theme running through most of them: the developer experience is being rebuilt around AI agents.

First up, a piece from Uno Platform called Developer AI Tooling in 2026: Trends Shaping How We Build. This is a wide-angle look at where the developer tool landscape sits right now. The article walks through five major trends. Agentic workflows have crossed from demos to genuine usefulness, powered not just by better models but by better scaffolding: structured tool schemas, planning prompts, and reliable execution environments. The terminal is having a renaissance with tools like Claude Code and GitHub Copilot CLI bringing AI directly into the command line. MCP, the Model Context Protocol, is quietly becoming critical infrastructure, acting like a universal adapter for connecting AI agents to databases, project trackers, design tools, and APIs. And then there’s the emergence of sub-agents and adversarial patterns, where specialized agents handle distinct roles and even argue with each other to improve output quality. The core message is that AI isn’t just autocomplete anymore — it’s the co-pilot, the reviewer, the tester, and increasingly the whole crew.

Next, a detailed walkthrough on Medium about building an end-to-end AI agent on Google Cloud Platform. The author built a football statistics assistant using Google’s Agent Development Kit, BigQuery connected through Cloud API Registry’s managed MCP server, and Gemini 2.5 Flash as the model. What’s notable is how little glue code is needed — the Cloud API Registry provides a managed MCP server for BigQuery, so the agent can query data via natural language without running any MCP infrastructure yourself. The article walks through two deployment paths: Cloud Run for a stateless REST API, and Vertex AI Agent Engine for a managed runtime with a playground UI.

Third, Seroter links to a Fireship video about Google Stitch, a new tool for UI and UX design. If you’ve ever struggled to turn a design mockup into working frontend code, this looks like it could be a lifesaver — part of the broader trend of AI collapsing the gap between design intent and functional implementation.

Fourth, a piece by Addy Osmani asking a provocative question: is the IDE dead? The answer isn’t a simple yes — it’s more nuanced. The center of developer work is shifting away from line-by-line editing toward supervising agents that can plan, rewrite files, run tests, and propose changes. Cursor just shipped Glass, an interface where agent management is the primary experience. Similar patterns are showing up in Claude Code Web, GitHub Copilot Agents, and tools like Conductor that let you run multiple agents in parallel isolated workspaces. The new loop is: specify intent, delegate, observe, review diffs, merge. But the article is careful to note that IDEs still compress genuinely hard problems into a high-fidelity feedback loop. So it’s less a death and more a center of gravity shift.

Fifth, a Google white paper titled Cloud Infrastructure in the Agent-Native Era. The argument is that cloud-native infrastructure was a step forward, but AI applications and agents need something further. If cloud-native was about containers and microservices, agent-native is about infrastructure that can support autonomous, tool-using AI systems with their own resource needs and governance requirements.

Sixth, a post about the three pillars of JavaScript bloat. This digs into why npm dependency trees keep getting bigger, identifying three culprits: packages supporting ancient runtimes and edge cases like cross-realm values, atomic architecture where every tiny utility gets its own package, and legacy code that simply hasn’t been updated. The e18e community has been doing good work pruning these, but the core insight is that a tiny group of people with niche compatibility needs are imposing costs on everyone else.

Seventh, a Forrester piece arguing that enterprise architecture is on the hook for generative AI success. GenAI doesn’t just add new tech; it increases architectural entropy unless deliberately integrated, governed, and continuously steered. Many organizations saw their proofs of concept proliferate, then hit second-order effects: stalled pilots, spiking inference costs, agentic systems crossing implicit authority boundaries. The remedy is shifting EA from static documents to living systems of knowledge, defining guardrails early, and making feedback loops non-negotiable.

Eighth, AI Doesn’t Fail in the Demo — It Fails the First Time You Have to Trust It from The CTO Advisor. This makes a sharp distinction between capability and control. We’ve largely solved for capability — models, performance, and tooling are good enough. What’s not good enough is our ability to control how these systems behave once deployed. The cloud analogy is apt: cloud succeeded because control became programmable through IAM, VPCs, and audit trails. AI doesn’t have an equivalent yet. The call is for separation: the system proposes an action, a policy layer evaluates it, and execution happens only if it passes.

Ninth, Robert Glazer’s Friday Forward on “Same Old”. The argument: very few things are truly unprecedented. AI disruption? The Industrial Revolution displaced millions. Shifting alliances? Britain and France were enemies for centuries before becoming partners. The antidote is studying history and understanding cognitive biases. The word “unprecedented” usually says more about the speaker’s knowledge of history than about the event itself.

Tenth, a TechCrunch story about Cursor admitting that its new Composer 2 model was built on top of Moonshot AI‘s Kimi, an open-source model from a Chinese company backed by Alibaba. An X user noticed code identifying Kimi as the underlying model, and Cursor initially hadn’t disclosed this. To their credit, they acknowledged it quickly — only about a quarter of the compute came from the Kimi base, with the rest from Cursor’s own training. The open-source ecosystem works precisely because people build on each other’s work. Just be transparent about it.

Finally, a deep dive from Slack’s engineering blog on rebuilding notifications. The legacy system had four conflicting mental models between desktop and mobile, hidden coupling between notification content and delivery, and inconsistent state across clients. The redesign introduced three clear channel options (all new posts, mentions, or mute), decoupled activity from push notifications, and achieved true cross-platform parity. A nice engineering detail: rather than migrating at the database level, they used a read-time strategy to reinterpret old preferences through new decoupled logic, preserving backward compatibility while enabling rollback.

The throughline today is unmistakable: AI agents are reshaping developer tools, infrastructure, and organizational architecture. But the excitement of what’s possible is running ahead of the hard work of governance, trust, and control.

Articles Covered

1. Developer AI Tooling in 2026: Trends Shaping How We Build — Uno Platform
2. End-to-End AI Agent on GCP: ADK, BigQuery MCP, Agent Engine, and Cloud Run — Medium / Google Cloud
3. Google just changed the future of UI/UX design… — Fireship (YouTube)
4. Is the IDE dead? — Addy Osmani
5. Cloud Infrastructure in the Agent-Native Era — Google (PDF)
6. The Three Pillars of JavaScript Bloat — 43081j
7. Architecture Is On The Hook For GenAI Success — Forrester
8. AI Doesn’t Fail in the Demo — It Fails the First Time You Have to Trust It — The CTO Advisor
9. Same Old (Friday Forward #528) — Robert Glazer
10. Cursor admits its new coding model was built on top of Moonshot AI’s Kimi — TechCrunch
11. How Slack Rebuilt Notifications — Slack Engineering