akatz

AI Art Appreciator, Vibe Lifer and Claude Abuser.
Building ComfyGit (among many other things).

[akatz @akatz_ai]

@Machinedelusion @TencentHunyuan Hell yeah! Does it work with one image or do you need multiple angles + video?

[akatz @akatz_ai]

@cerspense Sweet! The Depthflow stuff I did was all based of off @Tremeschine work so I want to make sure I credit him too 😁

[Spencer Sterling @cerspense]

Never been able to explore latent space like this before..
👀

[akatz @akatz_ai]

I love the ideas in the post and the article, but the full manifesto needs a serious edit. All the ideas could neatly fit in 1/4 the size. The number of times the same concepts with slightly different examples are explained is mind-numbing. Like how many times do we need to hear about 'Targeting Systems' and 'Blind Clears' and 'outcome based finances' etc.?

It seriously needs to take it's own prescribed medicine and be made more 'legible' with 'adversarial reads' to find where cuts can be made. Maybe AI can do it?

[akatz @akatz_ai]

@cerspense That's insane! Are you running Claude or an open source model like Qwen 3.5? I feel like this kind of workflow would be ideal for some OS models that can just chug away for days at a time

[akatz @akatz_ai]

https://t.co/jDCEm9LBkX

[akatz @akatz_ai]

I got Depthflow running entirely in the browser! 🌊

It uses an ONNX version of DepthAnything V2 tiny that loads and runs on the client, no server necessary.

Slap in any image, get a 3D parallax effect, and then export as webm, mp4, or GIF in 4K.

You can try it out for free here:
https://t.co/WnIRM5iBPL

Source code in the comments.

[akatz @akatz_ai]

One trick with agentic coding I find super OP:

Tell Claude to first run an 'experiment' to prove that the proposed approach will work before actually implementing any changes.

9/10 it will find some differences in its initial assumption. This is the "Trust but verify" strat.

[akatz @akatz_ai]

Strong planning up front with a mix of GPT 5.2 Pro on the webUI, multiple spec audit rounds with Opus 4.6 and Codex 5.3 xhigh, phased implementation with 5.3 xhigh doing the work with audit + manual validation after each phase, and finally multiple code reviews with full e2e tests covering expected behavior. Works for me so far.

[akatz @akatz_ai]

@JulienAIArt @Kling_ai Good! Been in the weeds working on a few things that might be useful to you, will hit you up on Discord about it soon😉 Hope you're doing well!

[akatz @akatz_ai]

@jboogx_creative like this but needs more tiny screens and RGB to compete with yours 😅

[akatz @akatz_ai]

I would say 'source available' instead of 'open source', here's a comparison of the repo license (PolyForm) vs MIT:

MIT is a permissive, standard open-source license allowing broad use (commercial, modification, redistribution) with only attribution; PolyForm Shield 1.0.0 is a "source-available" license that looks permissive but adds a significant competition clause, preventing users from using the code to compete with the original licensor's business, making it more restrictive than MIT, and generally not considered true Open Source by OSI standards, despite being similar in granting broad usage rights otherwise

[akatz @akatz_ai]

I’m so hyped for these models to get even better, the fact it’s OSS and runs on consumer GPU is amazing

[Eric is Daydreaming @ericxtang]

@DaydreamLiveAI + @overworld_ai = 🖖

[akatz @akatz_ai]

@gregpr07 Here's a simple agent orchestration 'framework' I'm building that's just TOML files and a Go binary: https://t.co/bhDPQSDUUF

[akatz @akatz_ai]

Everyone's got a new Multi-Agent strategy, so I may as well throw my hat into the ring:

MEOW (Meow Executors Orchestrate Work) is a simple scripting "language" that lets you build your own 'agent factories'.

The name is a nod to @Steve_Yegge 's incredible GasTown concept of the same name which inspired the effort (I'm a big fan of Beads!)

It's a single binary that orchestrates agents via
tmux. No Python, no cloud, just TOML files you can version control.

Most people already follow some kind of workflow when they're building software with LLMs:

1. Break a plan into small bite-sized chunks
2. Spin up one or more agents to tackle the work
3. Have each agent review their work
4. Ensure tests are passing
5. Commit changes
6. Merge the 1 (or more) worktrees back into main.

I was doing this for most of my projects already (manually) by watching each agent work and feeding it the next prompt in the 'chain' until it was done.

Now with MEOW I can crystallize ALL of my manual processes into automated pipelines:

- One workflow to unleash the swarm
- One for multi-phase review
- Another for QA testing and feedback

How does it work?

Define a workflow in TOML, run 'meow run my-workflow', and MEOW handles spawning agents, injecting prompts, waiting for completion, and coordinating dependencies.

Sort of like the Makefile of agentic coding.

The best part? You can SHARE these workflows with anyone.

I see so many 'good' ideas on my timeline for how to make agents write better code, or make the code more secure, less bugs, yada yada yada.

But we live in the 'one-click' install age.

I don't want to have to remember the 20 'best' ways to get agents just to do the damn thing:

"If you don't plan this way you're NGMI: First spin up 4 agents, have one be Aristotle, another Socrates, and another Linus Torvalds. Have them fight to the death over your PRD design. Finally have a 'moderator' agent pick out the best ideas. Hope you know how to use tmux!"

NO! Just give me ONE command that does all of that!
(Or give it to Claude to figure out)

And workflows don't even have to be that crazy.

Just want one agent that does a standard Plan -> Implement -> Test -> Review?

Spin up a workflow for it.

Want an entire software department with multiple hierarchies of programmers and managers and reviewers all working together to build out your company?

Hope you have tokens to burn, but yes.

The early Alpha version is up on my GitHub if you're insane enough to give it a shot (link below).

I'll try and post more 'example' workflows as I use them more in my own projects.

[akatz @akatz_ai]

https://t.co/1y6O1LabQQ

[akatz @akatz_ai]

Quick tip:

If you juggle multiple Claude Code accounts,

Download and install the excellent CAAM tool (Coding Agent Account Manager) by @doodlestein (github link below)

Run /login in Claude Code (CC) to login with your first account.

Exit CC and type 'caam backup claude <account1>'

Open CC and again run /login, but this time with your second account.

Again run 'caam backup claude <account2>'

Now you can switch between them whenever you run out of usage via:

caam activate claude my-account1

or

caam activate claude my-account2

You can do this with as many accounts as you own. It can also be used with codex, gemini etc.

Enjoy!

[akatz @akatz_ai]

Slowly but surely we are automating ourselves out of the need to automate ourselves

[akatz @akatz_ai]

@PurzBeats @nebsh83 This is straight 🔥🔥🔥