Happy Monday! Today's a great day to take that risk you've been waffling on. Go for it!

For big corporations (and people who care about consistency), I think we're going to see a push to decouple the versioning of harnesses, prompts, and models. It's one thing to get a smarter model, it's another thing to get a whole different harness experience.

Harness stability is going to be a real concern agentic tools going forward. I went from having a Claude Code that I could do mostly anything I need to a Claude Code that overwhelms me with walls of unneeded text. Can I go back to where were in the Opus 4.6 days?

Since when did Claude Code get so risk averse?!

I think it would be helpful if AI companies were a little more explicit when they were modifying their models and infra. Changing the prompts in agent harnesses is also an issue. Feels like I'm working with a different models.

The agentic era is gonna be huge for devops and infra companies. The need for virtualization, machine snapshots, infra as code, db cloning, you name it goes through the roof when you have a single human programmer managing even 5 agents working on a non-trivial system.

Also: who convinced Claude that answer questions with a 1000 word essay is the right move. It's like that friend who doesn't hit send while they respond and just dumps you with a full screen of text. This has also been getting worse of late.

Woke up. Started work. Today's Claude is a different Claude than a few days ago. I would really love a peek behind the scenes to know what changes when using Claude Code from having a highly productive Claude who gets my codebase and project to one that's doing the dumbest stuff.

If you're using a system that uses a database, using multiple agents with worktrees or separate checkouts do not ultimately solve the problem of colliding agents. You need full isolation.

Having now built a CLI, a stdin MCP server, and a remote MCP server, I would say that if your MCP server doesn't need to do anything locally, go the distance. Create the remote MCP server for this one reason: No versions.

It's amazing what one guy with a clear vision can do. I am that guy.

AI is an amazing paint brush and poor artist. AI is an amazing programmer and poor system designer. If you can get your thinking right, AI is a rocket shovel.

Thank you @jdxcode for mise, hk. 🙇 I appreciate all that you've brought to the world.

Even if AI could only fix my vim config and other setup, it would be worth it. Just unified my keymaps across Zed and neovim and boy is my muscle memory happy.

Those refactors you been putting off, Those doc updates you been deferring, time to enlist your favourite agent. A finely manicured projects is within your reach.

We pack all sorts of workflow efficiencies into our dotfiles. Agents can benefit from those tools too. The question is how do we setup a dev experience that's reproducible and agent-friendly. Same tools for humans and agents. Clear description in the AGENTS.md of what exists 🤔

AGENTS.md

AGENTS.md is a simple, open format for guiding coding agents. Think of it as a README for agents.

agents.md

I remember when I used to do day long rebases. Boy am I glad those days are over.

I think @mitchellh's advice to always have an agent running doing research on your next task is a great idea. I've been doing it and it's had a couple interesting effects.

New Claude verson. Back to it thinking it knows what I want my PRs to look like more than I and straight. @bcherny @_catwu did the env variable to turn off the git prompt get removed? I have it on and everything was 10/10 for a few releases.

If you're tightly coupled to a big platform like Shopify and they have a version release cycle and you have an API that needs versions, use the same one. It'll save you the headache of reasoning about what's valid when.

Did a demo today of a tool using a first-of-its-kind data store and seeing the agent make use of the data was really cool! 🤖

I create disposable tools to help me solve a problem and I've started leaning more and more on the @vuejs and @vite_js ecosystem. The agents is able to build complex things without a ton of setup or ceremony. React would be fine too. The key is its a highly reactive client side only framework.

--dangerously-skip-permissions is the only way to live. Why did it take me almost a year to adopt it? 🤦

Is it me or is GitHub degrading by the day?

Low-level programming can seem kinda scary when you've spent most of your time manipulating strings for web applications. What's interesting is if you give it a moment and squint a bit, it's all just blobs of data in loops with different properties.

I think it's a mistake to try to make a model do deterministic things like comply with a commit format or format code. Do you format code deterministically? No. You use your language's formatter.

Prediction: Once the software industry catches up with the use of agentic tools, we should see a net increase in the quality of "good" software. Teams that know how to create great software are going to leverage agents to make even better software.

Never would have thought it, but here I am designing a connection protocol. Here I am driving a @OmarchyLinux box via macOS Hardware encoded/decoded video Full clipboard sync Drag'n Drop file transfer from client to server Full audio sync "Time to teleport in"

The key to success with agentic is sometimes not accepting the first, second, or even third answer. Push the model.

What's fun about agentic development is that you can learn by building instead of learning by reading. Don't get me wrong, the study still has to happen, but it can be much easier because you can ask, "Give me a crash course on ...." Need more, keep asking for more.