[All writing](/)
Here's the June edition of interesting things I've read (and watched) this month, you can also [read last month's post here](/may-2026-reading-list).
[Retiring at the top](#retiring-at-the-top)
My favourite this month wasn't a blog post but a podcast: The Pragmatic Engineer's interview with Kelsey Hightower, Kubernetes and retiring at the top. Beyond the Kubernetes stories, what stuck with me was his perspective on AI - measured and grounded in decades of watching technology waves come and go, without the hype or the dismissal that dominates most of the conversation right now. There's a rare candour throughout the whole episode, from how he thinks about his career to what actually matters in this industry. Well worth the time even if you've never touched Kubernetes.
The Fable 5 rollercoaster June was dominated by the Claude Fable 5 story. Anthropic launched Claude Fable 5 and Mythos 5 early in the month, introducing a new "Mythos-class" tier above Opus. Simon Willison described the model as relentlessly proactive, which matches my early impressions - it takes a lot more initiative than previous models, for better and occasionally worse.
Then things took a turn: the US government issued an export control directive suspending all access to Fable 5 and Mythos 5 by foreign nationals, before Anthropic announced they were redeploying Fable 5 at the start of July after the controls were lifted, with updated cybersecurity safeguards. Quite a month for anyone trying to build products on top of these models - a reminder that model access is now subject to geopolitics, not just pricing and rate limits.
Amongst all of that, Claude Sonnet 5 was also released at the end of the month, and Anthropic announced Claude Science, an AI workbench for researchers that produces auditable artifacts.
Enthusiasts, sceptics and the human cost Charity Majors wrote that AI enthusiasts are in a race against time, AI skeptics are in a race against entropy. It's the best framing I've seen of the divide - both sides are responding to a real existential threat, and both feel like they're screaming into the void.
In a similar vein, AI Feedback Loops and the Human Cost of Failure looks at why volume doesn't equal productivity and why trust is harder to generate than code. The costs and benefits of AI adoption don't land on the same people, which goes a long way to explaining the rift.
The Pragmatic Engineer's slow down to speed up when working with AI agents covers the quality and tech debt problems from devs generating twice as much code as six months ago. The line that stuck with me: organisations that were dysfunctional already are now more dysfunctional - they're dysfunctional faster. It reinforces why I think measuring should come before tooling when rolling out AI across an organisation.
Joe wrote a short post on how we shape our tools, and they shape us back - with AI we're speedrunning that loop. And for the macro view, What Would It Look Like If the AI Bubble Popped? explores the contagion risks from the data centre investment boom.
Building with agents Addy Osmani published Loop Engineering, arguing that prompting skill matters less than being good at building the loop that does the prompting for you. Related in spirit is obra/superpowers, an agentic skills framework and development methodology that's worth mining for ideas even if you don't adopt it wholesale.
I also came across AgentsView, a local-first app for reviewing AI agent sessions. As more of our work happens through agents, tooling for actually inspecting what they did feels increasingly important.
On the platform side, the Enterprise-Managed Authorization extension to MCP is now stable, letting organisations provision MCP server access centrally through their identity provider instead of per-app OAuth dances. This removes one of the bigger friction points I've seen with rolling out MCP servers across an organisation.
Finally, two takes on design systems in the agent era: Impeccable packages design skills and anti-patterns for coding agents, and Meta released Astryx, an open source design system billed as "agent ready". It's good to see the industry moving towards making design systems consumable by agents, not just humans.
Interesting tech reads Why is DuckDB Fast? is a good deep dive into DuckDB's internals. I use DuckDB in usegraph, my open source project for scanning dependency usage across an organisation, so it was interesting to understand more about what's happening under the hood.
PlanetScale argue that the only scalable delete in Postgres is DROP TABLE. Large DELETEs add work instead of reclaiming it, so structure your data such that deletion becomes a DROP TABLE or TRUNCATE.
Pyodide 314.0 was released, and with it the acceptance of PEP 783 for Emscripten packaging - a big milestone for Python in the browser.
How LLMs Actually Work is a from-the-ground-up walkthrough from tokens to transformer blocks to the next-token loop. A good refresher if, like me, your mental model is fuzzier than you'd like given how much time we all spend using these things.
And Shopify wrote about Quick, their internal platform that lets anyone ship a site in seconds. The interesting part isn't the tech, it's the culture change - lowering the cost of sharing something running changes what people build.
Tools & opinions SQL to ER Diagram is a neat free tool - paste in CREATE TABLE statements and get an interactive ER diagram, entirely in the browser.
Ilograph published 7 More Common Mistakes in Architecture Diagrams, a follow-up to their earlier post. Finally, Stop Using Conventional Commits is an entertaining rant that makes some fair points about how the standard encourages focus on the wrong things. I don't fully agree - the structure is doing useful work in automated changelogs - but it's a good prompt to ask what your commit conventions are actually buying you.