Yesterday I asked Fable a stupid question: what if the memory-wiki I built for my agents became the harness itself?
~12h and an absurd swarm of subagents later, it works.
It can run agents, spawn subagents, use my providers through OAuth, preserve its brain through compaction and harvest every session back into its own memory.
Written in TS, obviously, the brain was already built that way.
First screenshot is the swarm that built it. Second is the harness actually running :D
Everything from a remote VPS
What a time to be alive!
So much info being gathered and it's only 14% of what we have to ingest!!!
Everything is knowledge, your agents needs an easy way to access knowledge.
This human view is just for fun and understand what's going on.
Turns out, after 2 days of debuggnig, it was the Bluetooth service.
As soon as I start it, the system start to lags, when I use my Echo as a speaker to play music (or even silence) it just constantly locked my system, something NUTS.
Anyone is experiencing Bluetooth locks on Linux with an AMD 9950X3D?
This is bananas. I guess no bluetooth for a while.
My computer has been crashing for two days while running agent work. It might be some lurking script but the whole PC crashes and needs a forced reboot.
I am tired, going to sleep. Tomorrow mayb I should format or something...
Ugh...
Fable is just amazing.
Themes actually work now
Goal is <1 % CPU usage BTW.
Without animations it eats < 0.1%
I don't know why GPT 5.6 SOL was unable to make this, I swear to god, all the styles were messed up, 39k lines of nonsense... And using their own harness!!
Me: Starts to develop his own harness
XAI: We open source our harness
FUCK
Should I fork and rewrite my harness in RUST or should I keep with what I have?
-_-
We've open-sourced Grok Build and have reset usage limits for all users.
Open sourcing Grok Build allows anyone to support making a reliable and robust harness. Check out our code, including the Git repo for the Grok Build CLI.
https://t.co/3SSvPu2Nrz
Google's got the chance to create the best generative AI models to exist, for image and video, yet they refuse to make anything useful or meaningful.
I use Gemini constantly for Vision analysis (it is really good) and to write documents, but that's it.
And even with my Google One sub, I hit constantly limits just by writing documents... IDK what game they playing but the wrong one.
@maria_rcks Agree.
I get that behavior is basically to avoid overloading the system or whatever, they might've found the key to allow all of us to consume without worrying.
But this could also mean they will eventually reduce the weekly limits, little by little.
Oh and it also supports GPU 3D rendering in the terminal and image rasterization but anyway, I know I have to give my images more detail, halftones are not enough.
But have you seen that 3D in terminal, anon!?
Our Engine V2 is working perfectly.
I'm focusing on our Designer now.
It has advanced so much, you have full mouse controls, interactive layer grouping (you move an item onto a container, gets grouped, aligned etc).
The UX for the designer is still WIP but the mechanics work beautifully.
You can even run it in EDITOR/LIVE mode to test out your UIs.
It allows you to save your UIs in pure Typescript, ready to be used. And also Load them (well this right now... no, I'm on it :P)
Fable is just amazing.
Themes actually work now
Goal is <1 % CPU usage BTW.
Without animations it eats < 0.1%
I don't know why GPT 5.6 SOL was unable to make this, I swear to god, all the styles were messed up, 39k lines of nonsense... And using their own harness!!
I hate so much going to sleep while my Claude Agents are running with bypass permissions only to wake up for a prompt 30 minutes after I went to sleep asking "Can I read from my own memory daddy uwu"
And wake up to 9 subagents with 600k uncached context, 0 tasks finished.
What is "Bypass Permissions" even for?!
And before you ask, all my agents are YOLO mode
I really hate this.
:( I even gave him Snapshots, mockups and detailed pages in images, before creating the specs Y.Y
It is better than Opus, that's 100% facts.
But IDK what did they do to Fable, that guy creates amazing UX.
I will keep trying, I had my old prompts and so far I've tested the new GPT models does not behave like old ones at all. So that might be one of the problems.
Samsung Dex + Moonlight / Native android linux emulator + Tmux = You can just work anywhere without losing battery.
This is my new setup for remote work right now.
It's a shame you can't root a Galaxy S11 Tab.