Fatih Arslan

Software Engineer at @planetscale.
Creator of vim-go. Design addict. Photographer.
Previously at @github, @digitalocean

[Fatih Arslan @fatih]

My latest post on control theory and feedback loops has just been published. I’ll start from scratch and gradually build up feedback loops that are self-healing and resilient, capable of scaling thousands of databases.

Check it out: https://t.co/khsqPD8WmT

[Fatih Arslan @fatih]

@kwuchu Mines is so weird. Dunno how to get it even: https://t.co/YZaNd9zYK6

[Fatih Arslan @fatih]

Best part about agents (or agent loops are), they are async and stateless. You kill them, come back 3/4 days later, they give you the illusion as if they are running for days. Super cool tech.

[Fatih Arslan @fatih]

First of all, thanks for all the suggestions. Thank you all. But, a lot of people pitched me a new harness or tool :) My problem is not how I use agents, and I know Agents can have their own sandboxes, or there are apps that do agent orchestration. I didn't ask about them! I'm already using Codex and Cursor's macOS apps (including Cloud agents, which I love. I think sandbox's are future,but more on that later). And I also tried herdr on my terminal.

My problem is with the feature `worktree` and switching branches. That's it. Apparently, people using worktrees always keep a `main` branch, which is synced, and use worktrees with 1:1 mapping to branches. Trying to do a 1:M mapping doesn't work and is not suggested. So I'll have to change my tools and expectations around that.

[Fatih Arslan @fatih]

Every time I need a new copy of a repo, I'm like "okay, let me try worktrees again." And each time, I regret it. Now I have a few worktrees and they're painful to use. They feel so unnatural to the git workflow. Is anyone here using worktrees daily and have found some tooling or workflow that makes them work?

[Fatih Arslan @fatih]

@msfeldstein https://t.co/nZkEvviCKW is a very good blog post. I don't know if they are up for contract, but worth nudging.

[Fatih Arslan @fatih]

@VladimirVivien I think Agent's don't like using jj (yet) and I need to tend an Agent.md for them

[Fatih Arslan @fatih]

@grkn Mantikli, benim de oyle ama yine de kontorl ben de olsun istiyorum :)

[Fatih Arslan @fatih]

@rbranson Wow 😮

[Fatih Arslan @fatih]

@austinnickpiel Of course and I love them :) The worktrees I created are via Cursor btw. My problem is only on my localhost, I just think the interaction with worktrees and git is not pleasant

[Fatih Arslan @fatih]

@coxley_dev That would be cool. Even if I don’t use them, they could be useful to see how others use it.

[Fatih Arslan @fatih]

@marcosnils @ulusalomer I believe I tried jj once, but because it's not in the training data at all of some models, they got quite confused. Maybe they are better nowadays?

[Fatih Arslan @fatih]

@grumpy_panda123 You mean this: https://t.co/cZ5sQCzbjY ?

[Fatih Arslan @fatih]

@perrito666 > I either do the dance if list then cd then nvim

That's what I do, and it's a pain :/

[Fatih Arslan @fatih]

En bariz uyuzlugu. Ayni anda iki branch'in iki farkli worktree'de checkout ediliyor olmamasi. Diyelim bir klasorde "main" var, worktree actin, diyelim orada "foo" var. Worktree klasorunde "main" gecemiyorsun mesela. Ve mevcut klasorundeyken worktree'ye gecemiyorsun. Keske daha UX friendly olsaydi, hic sevmedim.

[Fatih Arslan @fatih]

This is a really good talk from @samlambert on how to use Agents (e.g. @cursor_ai), to scale a database, safely. It shows how great agents can be if your infrastructure provider is capable of working together with Agents. A must watch for everyone https://t.co/njApfjFCik

[Fatih Arslan @fatih]

Reposting for attention. If you can help, do it, if not, it's worth amplifying. Share and repost it with others.

[Iheanyi Ekechukwu @kwuchu]

Long post incoming, but I'm desperate at this point and need the internet's help. (For the TLDR, check the replies for links and action items)

So my little sister Chizo has been fighting leukemia for the past 2.5 years. She's been in remission twice, once after a bone marrow transplant from our brother and then again after a CAR-T treatment after another relapse last year. She is now in her third bout of the fight, as the leukemia has returned yet again in December of last year.

She's spent the majority of this year in the hospital and they tried an alternative CAR-T treatment, which did not work this time. Afterwards, the doctors gave her two options: manage her pain and spend time with her family and loved ones until it's time to pass on, or to take a In order to address this third bout, they tried an alternative CAR-T treatment that did not work this time. The doctors gave her two options:

1) Manage her pain and spend the rest of her time here with her family and loved ones
2) Take a life-threatening chemotherapy treatment that can lead to heart/liver failure based on the intensity of it to bridge her to her next treatment, whether it's another CAR-T treatment or a bone marrow transplant if they find a donor.

She's a fighter, so she chose the chemotherapy treatment and is still fighting to this day. So I'm here to ask for y'alls help as she continues this fight.

[Fatih Arslan @fatih]

We have our balls guy @BenjDicken and the team. We’re in good hands.

[shadcn @shadcn]

If you’re building something, a company, a product or a team, go find your @leerob.

Go into debt if you have to.

[Fatih Arslan @fatih]

It's such a joy to work with a Fast model, especially for refactoring. When you use Composer 2.5 Fast for example, and then have to use Opus/GPT for other high reasoning work, they will feel like slog. Composer 3 when.

[Fatih Arslan @fatih]

@_heyglassy @PlanetScale Welcome to the team!

[Fatih Arslan @fatih]

The automation also has a secondary agent that closes the loop. When you make comments to the PR, such as pointing out errors or suggesting improvements, another agent reads them and pushes a new commit with the necessary fixes. This secondary agent doesn’t run automatically, it’s triggered only when comments are made on the PRs that the first agent has opened (it recognizes these PRs based on a GH label).

[Fatih Arslan @fatih]

Docs are the eyes and ears of Agents. But we're moving so fast that they are always outdated. So hence an outdated doc also confuses the models. But it's also hard keep up with updating them. Agents are just faster than humans. But docs are also still valuable, not for humans, but for agents.

So the fix is to use agents to help sweeping up outdated docs. In my case, I created a @cursor_ai automation, that also hooks into our Slack channels. It runs every 6 hours, checks past commits, messages made by our co-workers in the Slack channels, and then opens a PR with doc improvements and suggestions.