OpenAI just dropped AgentKit at their DevDay event, and it's basically their answer to making AI agent development way less painful. Think of it as Zapier, but specifically designed for building and orchestrating AI agents.
Until now, building AI agents meant juggling a bunch of different tools. Custom orchestration, manual evaluation pipelines, connector setup, prompt tuning, and weeks of frontend work. AgentKit bundles all of that into one unified platform.
The toolkit has four main pieces:
OpenAI just announced that Codex is now generally available, and they're shipping some really practical features that make it way more useful for engineering teams.
The big additions are a Slack integration, the Codex SDK, and better admin tools. But honestly, the SDK and GitHub Actions integration are what really caught my attention.
Following the vibe research, we now get vibe engineering from Simon Willison:
Vibe engineering establishes a clear distinction from vibe coding. It signals that this is a different, harder and more sophisticated way of working with AI tools to build production software.
“vibe research”. Like “vibe coding”, it’s performing a difficult technical task by relying on the model. I have a broad intuitive sense of what approaches are being tried, but I definitely don’t have a deep enough understanding to do this research unassisted. A real AI researcher would get a lot more out of the tool.
The basic loop of doing AI research with Codex (at least as an enthusiastic amateur) looks something like this:
- Codex makes a change to the training script and does three or four runs (this takes ~20 minutes overall)
- Based on the results, Codex suggests two or three things that you could try next
- I pick one of them (or very occasionally suggest my own idea) and return to (1).
Story: ⭐⭐⭐ | Narration: ⭐⭐⭐⭐⭐
Gene Kim and Steve Yegge’s audiobook on vibe coding has excellent narration, but I’d recommend checking out their YouTube channel instead. Most of the book’s examples were already covered in their promotional videos, and the high-level approach makes you wonder how long the advice will stay relevant in this fast-moving space.
The core framework they present, the Three Developer Loops, is genuinely useful though:
Welcome to my vibe-coded blog, built on GitHub Gists. I've been following Simon Willison's excellent blog for a few months, and this post in particular sparked something for me: My approach to running a link blog. A few months later, here we are.
Why not? I'm preparing for a talk at work about AI and going beyond "vibe-coding". I figured I'd make my notes and preparation public as part of the process.
This site was built for fun with a terminal-style theme; fitting, since I spend most of my time in a terminal as a DevOps engineer. I know the title "DevOps engineer" can be controversial, but it is what it is.
The co-founder of Instagram, Mike Krieger, which is now at Anthropic, mentioned that they use the Claude Agent SDK in a automated Slack channel when there is a new incident. It is name Claude Pager and it gathers all the context of the incident and can be poked to give out infromation when needed.
I have been waiting to implement something similar in our stack, once we are finally off the waiting list of the Datadog MCP waitlist...
Source: Big Technology Podcast
After the release from Gemini Cli, Claude Code is now offerings plugins. It seems taylor made for the Enterprise or teams, to quickly share configurations between team members. Excited to try it out.
Source: Claude Code plugins
Mitchell Hashimoto describes how he shipped Ghostty’s unobtrusive macOS update UI using agentic AI with deliberate human oversight. 
Define the shape of the solution first. Hashimoto chose Sparkle’s custom UI path and a titlebar accessory concept, then used an agent to prototype UI states, not to “build everything.” Humans kept control of scope and taste. 
I think it is one of the most powerful ways to create clarity of thinking for yourself, because clarity of thinking is, in my view, one of the most important things. And you only can really get that by forcing yourself to write down your ideas. And so I suppose one pertinent or timely piece of advice is, if you care about this, do not use an LLM to produce text for YouTube.
Because first of all, if you're putting your name on this thing, at least today, the LLM is like, it's pretty easy to tell when somebody does that. But more importantly, you are losing the opportunity to do the thinking yourself. And that thinking that happens when you write is one of the most important types of thinking for growth.