the only issues with this is if the providers dont like it
we opened this because we saw claude adding basically the same feature
but can people from the labs confirm this is ok to do? (opt'd-in)
Never do this.
The most efficient way of working with an agent is to get it to succeed at it's task with 0 further follow up prompts.
We have gotten their with our own orchestrator now, and the failure rate is very low. We have worked on and in this 100% since July 2025.
I still review at the end, and if all good, that is it. I also have no need to check the agents regularly, and hide them in the background and they only expose when done.
That means this compaction would be redundant work being done, and something I also turned off instantly as soon as I saw Claude implement it.
CANDLE CREW is OUT NOW ✅🎃
A scarecrow stole all the light. You've got one Halloween night, a music box and a crew of hopping little candles to bring it back 🕯️
My @TheSandboxGame Halloween Game Jam game, fully vibed with @threejs 😅
Play it here 👇
studio.sandbox.game/play/373…
CANDLE CREW: The Stolen Light 🕯️🎃
A scarecrow stole the light. A little Keeper and a crew of living candles are going to win it back.
My Halloween Game Jam entry for @TheSandboxGame, made 100% with @threejs inside Sandbox Studio. Out October 16th 👀 🔊
a strategy of hiring ADHD developers who's brains, trained on hunter gathering, seem to be evolved for multiagent development
I’m willing to take the bet on the other side if you’re willing to bet
I do play videos games
A lot
I’m an insane video game nerd
I also use these ai tools daily for work
I promise you the exponential growth is insane
Who knows what the future looks like but odds are.
anthropic 30k a day. 1,000 people.
spawn 7k a day. 8 people.
are we more AI pilled than anthropic?
Lisan al Gaib@scaling01Oct 13Europe
EuropeConnected via North America Android AppAccount-level information, not a live location or per-post device.
PR's to Anthropic's monorepo are exploding so they built their own git
Yeah that is awesome :)!!!
Did you talk about your techniques?
We got our own agent workspace setup which can kick off prompts and land them very easily, and also self healing infrastructure etc.
How many PRs does the commits represent? What about the 'size' of them?
Again ours also can be minor small changes, to insane huge features...and the prompts might be 1 sentence to a 40 min voice to text xD
I am trying to be careful not to worry about metrics, although they are important, if they become a 'measure' you can easily introduce bad incentives u know?
Last 7 days for me, over 1000 PRs
Our Agent Workspace really makes it easier and easier each day!
here's how i shipped 2,500 PRs last month to production
this was originally supposed to be for Cursor Compile in London. i couldn't make it since i was livestreaming for Grok @Bot Galaxy so i'm making it available for free here on X! watch it on 2x speed, i talk slowly
here's how i shipped 2,500 PRs last month to production
this was originally supposed to be for Cursor Compile in London. i couldn't make it since i was livestreaming for Grok @Bot Galaxy so i'm making it available for free here on X! watch it on 2x speed, i talk slowly
the $500 plan from OpenAI is a total joke now
(and yes, i tested it myself)
The $200 is still on the old 20x for now, but will lose it soon. Unfortunately will likely need to upgrade end of month ~ when it happens.
Ultra fast is likely over-rated except for very specific workflows, but 99% TPS is not the bottleneck.
Hey so whats the deal with making games on Spawn? I keep seeing people post about it. Do you have to use their tools or can you make games in unity/godot and just publish to spawn?
Replying to @metatransformr @AIOnlyDeveloper
you can bring your own agent. export your game’s code. it’s your game.
there's effectively zero limitations and any you find we will fix and remove the same day
The biggest problem with AI is that it is capable but slow. And if you want it fast, you have to pay the big bucks.
I don't want to hand over $500 a month to OpenAI for UltraFast.
I just want an AI that can actually do what I said, faster than me... so there's a benefit.
The AI TPS is not the bottleneck, when you can run multiple in parallel, humans still are the bottleneck.
The $500 plan also is really not good, the Ultra fast will burn your credits in a few decently sized convos, only for PURE emergencies OR really important things where near instant feedback is needed.

Denys Cherednychenko@Denche2005Oct 7
Spain!
SpainConnected via Spain Android AppLocation may be affected by a proxy or VPN.Account-level information, not a live location or per-post device.
Hot take:
You can vibe-code a small app.
You can't vibe-maintain a 90K+ line codebase. No chance.

Boris Cherny@bchernyOct 6
United States
United StatesConnected via United States App StoreAccount-level information, not a live location or per-post device.
I am surprised that people are surprised this is how I prompt Claude.
Talk to Claude the way you would a coworker. There's no secret to prompting. There's no need to be overly scaffolded or prescriptive for most tasks -- give Claude a goal, and it will figure it out.
Back in the Sonnet 3.5 days, your prompt mattered a lot. Nowadays, it's much more important to communicate to the model:
1. What you want it to do
2. How much effort you want it to spend
3. How it should verify that it did the right thing
Boris Cherny@bchernyOct 6United States
United StatesConnected via United States App StoreAccount-level information, not a live location or per-post device.
Replying to @bcherny
Prompt

Boris Cherny@bchernyOct 6
United States
United StatesConnected via United States App StoreAccount-level information, not a live location or per-post device.
Another example, my actual prompts xitter.cf/bcherny/status/2102543…
Boris Cherny@bchernySep 22United States
United StatesConnected via United States App StoreAccount-level information, not a live location or per-post device.
I used Opus 5.5 to formally verify the Claude Agent SDK using Lean. A couple short prompts = 16 PRs fixing various bugs and race conditions. Video attached.
TLA+ also works well. I sometimes combine Lean and TLA+ to look for issues around data flow, concurrency, and state mgmt.
I don't know either language well, but Claude is excellent at both. This approach is super useful for formally modeling your code and finding bugs that a human probably wouldn't have spotted.
Is formal verification the future of coding (or at least, bug finding)?

Tibo@thsottiauxOct 3
United States
United StatesConnected via WebAccount-level information, not a live location or per-post device.
What’s one thing that’s missing in codex that you wish we had?

Tibo@thsottiauxOct 5
United States
United StatesConnected via WebAccount-level information, not a live location or per-post device.
Day 1/
We have optimized the default speed to be ~50% faster across GPT-6 Astra and GPT-6.1 Sol through the subscription across all our products and partners using Sign in With ChatGPT (including OpenCode, Pi, Amp, Devin, ...).
No changes needed on your end and this should be felt within the next two hours.
Tibo@thsottiauxOct 4United States
United StatesConnected via WebAccount-level information, not a live location or per-post device.
Over the next 28 days, each day we’ll either ship one thing that is a clear improvement and relevant for most codex/work users or ship a full reset. Let the improvements begin.
The Bottleneck is no longer TPS for myself, and I imagine various power users. Please offer a /Slow mode for those that are happy to go slower (as long as no cache issues that result in more costs), so that we can get more usage :D! Even happy to delay the 'start' of the agent also, a bit like 'batch' mode, although yeah might be a bit more thought needed.

Theo - t3.gg@theoSep 23
United States
United StatesConnected via United States App StoreAccount-level information, not a live location or per-post device.
I'm going to be real with y'all for a minute. I am so, so tired of voice-to-text.
I like typing. I can communicate exactly what I'm thinking without having to reserve mental energy to get the thing out of my head. The keys I press are the letters that appear on my screen. I can type quickly. I can get my ideas out quickly. It's great.
Because of my hand injuries and my cast, I am not able to type. This has forced me to become a power user of voice-to-text tools. They're not as bad as I expected. It was even a bit pleasant once I got the hang of it. But the more I use them, the closer I get to throwing my laptop out a window.
When I screw up typing, I get like 1-2 letters wrong, and usually you can tell what was meant. Fixing the mistakes only takes a few seconds.
When Wispr Flow fails to translate what I said, it removes and swaps random words in my statement, turning it into jumbled nonsense.
It regularly removes keywords from sentences, like "not", "never", and "before".
Do you know how big a gap there is between "never merge a change like this" and "merge a change like this"? Do you know how frustrating and embarrassing it is when actual damage is done by a Wispr Flow hallucination? Do you know how tiring it is to reread everything I just said, right after I said it, carefully checking that every word was transcribed correctly? Do you know how miserable it is having zero confidence that the thing you're trying to say will actually come out correct?
It's so fatiguing. It is making computers miserable. I don't know how people live like this. I hate it. I want nothing more than to be able to type again.
Given the choice between the ability to type and the ability to speak, I would legitimately be willing to give up my voice if it meant I never had to feel this way again.
I'm just now realizing how much more irritable, frustrated, and miserable I have been for the past month because I can't use a computer in a way that is pleasant. I legitimately feel the sanity leaking out of my head every time I have to hit the Wispr Flow button.
To everybody I've been ghosting or slow to respond to lately, I'm sorry. Know that I want to, but my inability to use a computer is driving me mad, and I don't want to respond when I feel like this. I'd much prefer to wait and give you the response you deserve rather than a pissed-off pile of slop like I would be able to do today.
I had to make 12 corrections to the following tweet before I could post. Making those corrections took significantly longer than typing would have taken (if my hand worked).
Damn, it's been a busy September so far, and there are still 8 days left! I think I can probably do at least another 20k commits in that time. I'm really flying now with Opus 5.5 and GPT-6 Sol. You should increase your own work expectations every time we get new models like this!
Version 2 😂
Took the CUDA-generated city experiment way further. STRATUM Infinite generates the city continuously as you travel, with the same procedural building grammar used from skyline scale down to windows, balconies, roofs and surface detail.
No asset streaming. No giant world file. Just .cu (to WGSL) generating the city as you explore.
samg-coder.github.io/stratum…
Takes ages to Compile...
github.com/SamG-Coder/stratu…
Been experimenting with generating assets on the GPU instead of streaming them.
Built a 2.3km procedural city where the buildings, geometry, materials and tiny surface details are generated from CUDA .cu (Precompiled*) as you explore.
Basically compute streaming instead of asset streaming 😂
github.com/SamG-Coder/stratu…





























