Seroter's Daily Reading — #785 (May 15, 2026) — Seroter's Daily Reading

Listen: https://blossom.nostr.xyz/3c562d9bab93d7eac9ade3309fdd9d8207b6125756e6aa02ad7c580a25fb8dcc.mpga

Seroter's Daily Reading, episode 785, May 15, 2026. No-meeting Friday, email backlog cleared, RSS reader emptied, writing done. Enjoy your weekend.

Let's get into it. We've got a packed episode today covering leadership theory, AI's reshaping of developer identity, the changing economics of AI subscriptions, and the architecture of AI agentic applications. Let's go.

First up, a piece from Harvard Business Review on Are You Meeting the Needs of the People You Lead? This piece is anchored in research across 3,500 workers in the US, UK, and China, and it identified six fundamental needs that followers consistently evaluate their leaders against. Protection, fairness, vision, expertise, affiliation, and status. The key insight here is that leadership failure is rarely about bad intentions or incompetence. It's about misalignment. A leader who excels at providing vision may fail spectacularly when followers need protection or fairness. The leadership style that works in one moment can be completely wrong in another. Nadella at Microsoft is the classic example: he didn't double down on performance pressure when morale was low. He rebuilt collaboration and restored a sense of shared purpose, addressing the affiliation and expertise gaps that had developed. The practical diagnostic the article offers is straightforward: leaders should ask themselves which follower need is most salient right now, identify their own default bias, and rebalance deliberately. What's powerful about this framing is that it moves leadership from "what kind of leader am I" to "what is actually needed here." That's a useful reframe for anyone in a management role.

Next, a piece from Research-Driven Engineering Leadership on How does deepening AI fluency change what it means to be a software developer? This is based on research combining surveys of hundreds of developers with in-depth interviews of AI-fluent practitioners. The core finding is that developers don't want AI everywhere. They have a clear sense of which parts of their work carry identity weight and which parts are just toil, and they're deliberately protecting the parts that make the work feel meaningful. Mentoring, relational coordination, and system design landed in the "keep this human" zone, not because they're easy, but because developers recognize that's where growth and authenticity live. For core work like coding and debugging, the appetite for AI help is high, but the blocker is trust. Current tools fall short on reliability, transparency, and contextual alignment. That's a tractable engineering problem, not a cultural resistance problem. The piece maps out four stages of AI fluency: Skeptic, Explorer, Collaborator, and Strategist. The Strategist stage is interesting: developers who reach it practice their craft at a higher level of abstraction, organizing around vision, direction, and verification. The highest-value skill becomes knowing what to build, why it matters, and whether what was built is right. Even optimistic adopters expressed worry about the learning paradox: are they getting better at engineering, or just better at prompting? And there's a real concern about how junior engineers develop without the boilerplate work that used to be their entry point. The practical guidance for engineering leaders is to map the AI opportunity space before scaling adoption, invest in trust-building for core work, and protect the on-ramp for junior engineers.

Speaking of AI tooling, let's talk about Anthropic and subscription pricing. InfoWorld reports on Anthropic puts Claude agents on a meter across its subscriptions, with Anthropic separating third-party agent usage from Claude subscription pools. Starting next month, running agents through tools like OpenClaw or the Agent SDK will draw from separate billing, not the shared subscription. This follows the earlier move to remove third-party tool access from subscription limits entirely. Developers are pushing back, arguing this undermines one of Claude's biggest advantages for agentic workflows: predictable subscription pricing for large-scale automations. The article notes OpenAI is apparently trying to capitalize on developer frustration here. This is a signal worth watching. When providers start metering agent usage differently from interactive usage, it changes the economics of building agentic systems. Subscriptions made long-running autonomous tasks budgetable. API pricing introduces variable cost that has to be managed. Whether this is a capacity constraint play, a pricing model evolution, or something else, it'll shape how teams architect their agentic workloads.

Shifting to architecture, the Google Developer Blog announced Announcing Genkit Middleware: Intercept, extend, and harden your agentic apps, a composable system for intercepting and extending agentic applications. Genkit is Google's open-source framework for building AI-powered apps in TypeScript, Go, Dart, and Python. The middleware hooks attach at three layers of the tool execution loop: Generate for context injection and conversation-level logic, Model for retry, fallback, and caching, and Tool for human-in-the-loop approval and sandboxing. The pre-built middlewares include Retry with exponential backoff and jitter, Fallback for switching models on failure, Tool approval for gating destructive actions, Skills for injecting SKILL.md files into prompts, and Filesystem for scoped file access. The real power, though, is in custom middleware. The blog shows a content filter written in about twenty lines of code that rejects responses containing forbidden terms. You can compose and stack middleware, with ordering matters: the first listed wraps everything inside it. The Developer UI ties it all together, letting you inspect and debug middleware execution trace by trace.

Following that thread, a DZone piece compares Vercel AI SDK Middleware vs Genkit Middleware: A Hands-On Comparison head to head. These two frameworks take very different approaches. Vercel wraps the language model itself, and the middleware travels with the model transparently to the rest of your code. It's configured once at startup and is static. Genkit takes the opposite stance: you pass middleware as part of each generate call via a use array, and composition is dynamic, changing per request based on user, tenant, or experiment. Vercel's built-ins smooth over provider differences: reasoning extraction, JSON cleanup, simulated streaming. Genkit's built-ins are about production hardening: retry, fallback, tool approval, skills, filesystem. The most important architectural difference is tool handling. Vercel's middleware sits at the model level, and agentic loops are handled by higher-level abstractions. Genkit has first-class hooks for tool execution and the agent loop itself. If you're building agents with significant tool calling, Genkit's surface is more expressive. If you're already deep in the Vercel ecosystem and need provider normalization, Vercel AI SDK is the natural fit. The article notes you might use both: Vercel on the frontend for chat UX, Genkit on the backend for orchestration and agentic flows.

Finally, a piece from The New Stack on How to build a skills library for your engineering team. This is a practical, operationally grounded piece. The problem: every engineer on a team runs a different version of their AI coding assistant, and skills are local config files, so there's no visibility or consistency. The solution is a centralized skills library. Skills are just Markdown files that tell AI agents how your company works: security conventions, incident protocols, coding standards. The steps: put skills in version control, organize them into groups, mark some as required and others as optional, and automate distribution so engineers get updated skills automatically. Required skills are non-negotiable and include security monitoring and coding conventions. Optional skills are role-specific, like a Django backend skill that activates for Python files. The piece describes a meta-skill that watches for repeated corrections and proactively suggests creating new skills. When you fix the agent on the same thing twice, it asks if you want it to write a skill. That self-closing loop is elegant: the agent surfaces its own gaps and routes fixes into the review and distribution flow. There's also a dashboard for tracking which teams have which skills loaded and flagging anything that's gone stale. The author frames this as solving an agent sprawl problem. Most companies already have it: hundreds of unofficial different skills and MCP servers scattered across the org. A skills library gives the platform team visibility into what every agent in the org knows right now. That's a capability most organizations currently lack.

That's episode 785. A few threads connect across these pieces. The HBR piece on leadership is really about alignment and context-awareness, which maps to something in the AI fluency research: developers aren't resisting AI, they're making rational bets about what to protect. Both pieces are about diagnosis before action. The Anthropic pricing change and the Google-Wiz integration both signal that the infrastructure layer is maturing and getting metered accordingly. And the middleware comparison and skills library pieces both speak to the operational reality of running agentic systems at scale: you need composable, observable, governable tooling. The skills library in particular is a reminder that the hardest part of agent adoption isn't the technology, it's the organizational plumbing that makes it sustainable. Have a good weekend.

Are You Meeting the Needs of the People You Lead? — Harvard Business Review
How does deepening AI fluency change what it means to be a software developer? — Research-Driven Engineering Leadership
Cloud CISO Perspectives: How Google + Wiz changes multicloud strategy for CISOs — Google Cloud
Is Software Losing Its Head? — a16z
Anthropic puts Claude agents on a meter across its subscriptions — InfoWorld
Announcing Genkit Middleware: Intercept, extend, and harden your agentic apps — Google Developer Blog
Vercel AI SDK Middleware vs Genkit Middleware: A Hands-On Comparison — DZone
How to build a skills library for your engineering team — The New Stack