Seroter's Daily Reading — #840 (August 6, 2026)
Follow into
Save into
Source: Seroter's Original Post
Seroter's Daily Reading, episode 840, August 6th, 2026. Ten articles today, and they're loosely organized around a theme: the tooling and infrastructure that AI coding agents run on is evolving rapidly, and everyone is trying to get the packaging, distribution, and evaluation of these capabilities right.
Let's start with two pieces from Google. First up is a post on the Google Developers Blog about Enable on-demand expertise with Agent Skills in Genkit Go. Genkit is Google's AI SDK for Go, and it's been getting some well-deserved attention. The post digs into how you can use Agent Skills with Genkit to package specialized expertise into discoverable, on-demand capabilities. The core idea is called progressive disclosure: you load only the metadata of a skill initially, just the name and description, and then the agent decides when it actually needs the full instructions. When it does need them, it loads the full SKILL.md body plus any bundled scripts and references. This keeps token usage lean early on, then fires only when relevant. The post has working code examples, including a pretty fun art restoration app that can look at a painting, recognize it as a painting, load the right restoration skill, and output a restored image. Worth clicking through just for the before and after on that Ecce Homo example.
The second Google piece is about Agent Plugins package your skills, tools, and more. This is an open, vendor-neutral specification for packaging Agent Skills and MCP servers together into portable bundles. Google is joining the core maintainer group alongside Amazon, Cursor, Microsoft, OpenAI, and Vercel. The spec is deliberately minimal: a plugin is just a directory with a plugin.json manifest, a skills folder, and an mcp.json for MCP server declarations. The insight here is that the fragmentation problem isn't the components themselves, it's the wrapper around them. Every client was inventing its own format for how to bundle skills with their scripts and MCP servers. Now there's one portable format that works across clients. The post also makes the point that not every skill needs to be a plugin. If you're shipping a single skill to a single client, just ship the skill. Plugins earn their keep when you have components that genuinely belong together and need to travel together. Good nuance.
Over to The New Stack for a piece on Today's Codex will feel "primitive" by fall — and its own team's roadmap backs it up. The head of core products at OpenAI posted on X that today's version of Codex will feel primitive in two to three months, and that the next generation of models needs more than a laptop. That last part is the interesting signal. OpenAI is in the process of acquiring Ona, which used to be called Gitpod, a company that creates secure cloud development environments. The idea is to give Codex a persistent workspace in a customer's cloud so that agents can keep working even after the laptop that started the task is closed. Agents that run for 25 hours straight and generate 30,000 lines of code are impressive, but they currently depend on the developer's machine staying online. Ona solves that. It also raises the stakes on security, which the article covers honestly: an agent with credentials to your cloud is a new kind of surface area. OpenAI knows this and is building customer-controlled access rules and activity logging. Worth watching as this acquisition moves through regulatory review.
Then there's a practical question that comes up constantly: Should You Self-Host Inference? The AI Engineer has a thorough breakdown. The short version: API is the default and it's cheaper under about a million tokens a day. Self-hosting wins on volume and control, with the crossover around two million tokens a day. Above ten million a day, owned hardware typically pays back in six to twelve months. But the article is honest about the hidden costs. The MLOps engineer to keep those GPUs alive runs about $160,000 a year in salary alone, which is more than the hardware. And the strongest models, the frontier ones from OpenAI, Anthropic, and Google, are closed-weight and API-only, so self-hosting means open-weight models like Qwen and Llama. The hybrid architecture is where most teams end up: sensitive and high-volume work local, hard reasoning going to the frontier API over the wire. That split typically runs 40 to 70 percent cheaper than an all-API stack.
Amanda Fitch has a fun piece on building a website entirely with AI agents, A new way to website, and it's worth reading as a lived experiment rather than a tutorial. She assembled a toolkit with Google Antigravity as the primary agentic platform, Claude Code as a secondary architect and security agent, and some Gemini models for image generation. The site started as a basic HTML blog, then she had the agents build a markdown-to-HTML pipeline, then she went down a rabbit hole on web design for 2027 and ended up implementing something the article calls the Oracle interface, which replaces traditional navigation with a conversational search that routes you to the right page. Agents can be stubborn about deviating from conventional layouts, so this took some back and forth. The bigger lesson is architectural: agentic programming moves fast but the code organization often suffers, so she spent time refactoring shared code into reusable libraries. The security pass also caught real issues: a third-party JavaScript library pulling fonts from servers she hadn't opted into, and unsanitized markdown-to-HTML conversion. She had her security agent patch those. The final workflow is just write in markdown, run one command, and the live site updates.
Google has another piece this week on Scaling AI Agent Infrastructure with the MCP Stateless updates. This is the team that led the charge to decouple the Model Context Protocol from stateful transport constraints. The original MCP spec required a session ID that pinned clients to specific containers, which broke horizontal scaling behind load balancers. The new 2026-07-28 specification removes the session handshake entirely, makes every request self-describing, and introduces HTTP headers like Mcp-Method and Mcp-Protocol-Version so that proxies and load balancers can route traffic without inspecting the JSON-RPC body. This is the biggest change to the MCP spec since launch, and it makes load balancing boring, autoscaling seamless, and serverless deployment actually viable for MCP servers. The GitHub MCP Server has already upgraded and removed Redis session storage entirely. Multi-Round-Trip Requests and a Tasks Extension for long-running async operations are also part of this release, which the working group built together with Hugging Face and others.
Meta AI Research released Muse Code and Muse Spark 1.2. The article wasn't fully fetched, but the description indicates Meta is getting into the coding CLI space with a strong feature set and early benchmark numbers. Everybody seems to be building one of these now, which suggests the agentic coding tooling market is getting crowded fast.
PostHog has a piece on writing agent skills, What nobody tells you about writing agent skills that is packed with battle-tested knowledge. The core idea is that skills are a superset of code, not a replacement for it. A skill should be precise about the goal and constraints and context the agent can't derive on its own, but ambiguous about the steps, the failures, and the runtime specifics. Over-specifying turns a skill into a brittle workflow. PostHog has 226 skills in their internal store and 187 SKILL.md files across 28 products, so they've learned this at scale. Another key principle: skills rot, because the products, APIs, and models they depend on change. Their answer is to split durable structure from volatile content, point skills at a single source of truth like documentation, and regenerate rather than patch when you do a fix. They even have a context mill pipeline that automates skill generation. The article also emphasizes that not everything deserves a skill: write them for work you do repeatedly, work that agents do badly by default, and work that can run on autopilot.
Google has a third piece this week, this one on Agent and Model Evaluations in Gemini Enterprise Agent Platform are now GA. The platform gives you pre-built metrics, adaptive rubrics that generate case-specific grading criteria, experiments you can run locally or server-side, and online monitors that grade live production traffic and produce drift alerts. The case generation component can bootstrap eval datasets from your agent's instructions, a user simulator plays out multi-turn conversations, and an environment simulator stands in for failing or slow backends so you can test agent resilience without touching production. The issue clustering feature groups eval failures into interpretable clusters against your own taxonomy. This is a full-featured eval loop that integrates with the Agent Platform SDK, agents-cli, and ADK.
Finally, Simon Willison on Technical Blogging. This is a short piece, essentially a pointer to an interview where he repeats his best advice: lower your standards and hit publish while you're still actively unhappy with what you've written. The only alternative is a folder full of drafts and never publishing anything at all. The flaws you see in your own writing are invisible to everyone else. Simple advice, hard to follow, and it's served him well for years.
That's episode 840. A loose theme running through today's batch: the infrastructure layer, from skill packaging and protocol specs to eval frameworks and inference deployment, is where a lot of the real engineering energy is going. It's not glamorous work, but it's the kind of unglamorous infrastructure that should be shared rather than reinvented five times over.
- Enable on-demand expertise with Agent Skills in Genkit Go — Google Developers Blog
- Today's Codex will feel "primitive" by fall — and its own team's roadmap backs it up — The New Stack
- Agent Plugins package your skills, tools, and more — Google Developers Blog
- Should You Self-Host Inference? — The AI Engineer
- A new way to website — Amanda Fitch
- Scaling AI Agent Infrastructure with the MCP Stateless updates — Google Developers Blog
- Introducing Muse Code and Muse Spark 1.2 — Meta AI Research
- What nobody tells you about writing agent skills — PostHog
- Agent and Model Evaluations in Gemini Enterprise Agent Platform are now GA — Google Developers Blog
- Simon Willison on Technical Blogging — Simon Willison