Johan Carlin

Corollary to this - biggest mistake I see people do creating a MCP server is reflect their entire REST API shape as MCP tools.

You don’t have to do that. 10 REST API endpoints for placing an order can be easily made into one tool - place_order. Hide the wiring behind the wall.

If however you don't have a clear vision for what this thing should become, the agents start spinning their wheels in no time. It's great to be able to test your ideas like this. Sometimes it doesn't come together, and now the fault is usually with the vision more than the agent's ability to execute

Sure, some of the orchestration of the plan-build-PR pipe could probably be delegated too. But there is something you have that the agent does not and that is *intent*. In this sense the situation with agents is not so different from working with skilled but poorly business-aligned engineers.

Take a capable agent like CC and it's even worse. You put down your half-baked shower thoughts and ask it to make a plan. It's beautiful. You tell it to build. It does, wonderfully. Another agent looks at the PR and you wave it through. Could you go gas town and replace yourself with another agent?

Leading engineers is bad enough. Half the time you feel badly out of your depth. You handwave at a technical approach and hint at the business value and good engineers take that and build something amazing. It's easy to feel like an imposter but I've seen what a difference that direction makes ...

Ok great so terminal wins for agents but then the question becomes how do you sandbox it effectively because the current YOLO approach is going to become a security disaster if widely adopted

screenshot of a claude conversation that says too many em-dashes You're right--let me fix that:

🤨

One of the problems with being UTC+1 is that my morning bluesky feed is mostly late night, bleary-eyed, two-drinks-in American expressions of rage and grief over the death of the republic. It's quite the way to wake up

Surprised to see OpenAI disable SSL validation here in the chatgpt sandbox. I'm sure it's a manageable risk in their setup but kind of unnecessary to accept it

A very sane AI usage policy for any open source project that still cares about quality.

One of the hardest thing for humans to do consistently is get risk factors in the correct order. Confusing primary effects with third order fourth order or fifth order effects is shockingly human.

This kid is now over 20… I wish some reporter would do a follow up story

The Lidl panettone has been open for a month now and it's still completely fresh. A few more Christmases and when it's time I won't need embalming

Codex Web is similarly handicapped at the moment, presumably it's a security thing (the repo branch is checked out before the agent starts so it has no access to credentials)

the rate of change in the AI space is so fast that i cannot explain to my work colleagues that the API we have access to (GPT-4o) is so outdated as to be useless because there's no single chart that depicts both models on the same benchmarks.

Spent stolen moments during the holidays vibing a refresh of my old blog with codex web. I came away with a fancy Sveltia CMS and an appreciation for Codex Web, which is limited yet highly addictive www.johancarlin.com/vibing-my-bl...

docx linters anyone?

There's a step down in agent performance from software development to standard office apps. In part it's that the tooling is not there yet (office/gdocs gui integrations are crude compared to VS Code).

But in part there's just no feedback loop without automatic tests and that's harder to fix

Creator of node.js

Brian Eno: "The great benefit of computer sequencers is that they remove the issue of skill, and replace it with the issue of judgement."

Spent a bit of time vibing from my phone with codex web. The big showstopper so far is it has no way to handle merge conflicts or indeed respond to any incoming changes from the feature branch. Makes it tricky to run parallel tasks

Recruiting developers is tricky these days, half the CVs are LLM-reworded versions of your job ad...

Gas Town Emergency User Manual It’s been a busy 12 days since I launched Gas Town. ️I’ve merged over 100 PRs from nearly 50 contributors, adding 44k lines of code that no…

WTAF
steve-yegge.medium.com/gas-town-eme...

The future is unevenly distributed.

We really need to come up with better metaphors for AI agents than junior developers and interns. It’s far more of a reflection of how disrespectfully we treat those roles than it is a useful metaphor for how agents behave

Yeah it's exactly backwards. In our codebase we're having to write documentation because the LLM doesn't do word of mouth. Onboarding LLMs means writing everything down in the repo and that's a huge win for humans also

Stateful agentic LLMs are cool and all but I hope we get past the anthropomorphizing tamagotchi phase soon. I don't think it's a helpful mental model to think of these things as individuals with agency

Has nobody dealt with their legal, finance, and procurement folks at their companies? Y'all realize that people buy software to have "a throat to choke", right?

The functionality is... not that important. It's why RedHat exists, for example: the software is free, but big companies don't *want* free

Why Opinionated, Code-First Systems Are the Future of Data How AI amplifies strong abstractions instead of replacing them

The idea that code-first BI tools will gain traction so that AI workflows can generate end to end analytics pipelines that work with both the opinions of the tools and the opinions on the developers is really compelling.

open.substack.com/pub/hanson37...

Tibo @thsottiaux X.com Background Terminal is a Codex feature we've been baking for a bit and will soon turn on by default. More powerful than it sounds, it allows to Codex to multiplex work and significantly speeds things up. Default-on within OpenAl. You can enable with / experimental in the CLI. 10:01PM • 12/27/25 • 87K Views 63 17 59 737 203 Relevant v View quotes > Tibo @thsottiaux• 19h ... When the feature is enabled, you can often catch the model doing something else while it waits for a command to finish for example. Or spin up multiple ones to run multiple test suites concurrently. 11 t7 13 • 204 ıll 8.2K 企

background terminal in Codex seems like it would really speed things up

The occasional Google inactive account manager reminder emails are a real memento mori