If you're curious how I managed to do over $20,000 in inference on the last 48 hours, here's a video all about it.
Spoiler: loops are really powerful
7:28 PM · Jun 18, 2026 · 54K Views
Sentiment
Positive users praise coding agent loops as a powerful cheap-iteration 'cheat code', while negative users resent the $20k inference costs as unaffordable and anxiety-inducing.
Pos
75.0%
Neg
25.0%
21 comments with sentiment.
Cluster Engagement
Digg Deeper
No Digg Deeper questions have been answered for this story yet.
Posts from X
Most Activity
Most Activity
VIEWS17.2KBOOKMARKS35LIKES163RETWEETS1REPLIES13
Theo - t3.gg@theo
One of these loops went so bad that I made Codex write an apology document
Theo - t3.gg@theo
If you're curious how I managed to do over $20,000 in inference on the last 48 hours, here's a video all about it.
Spoiler: loops are really powerful
2hViews 17.2KLikes 163Bookmarks 35
Theo - t3.gg@theo
I said “less emdashes and more guilt” and it added this section 💀💀
Theo - t3.gg@theo
One of these loops went so bad that I made Codex write an apology document
2hViews 3.9KLikes 29Bookmarks 3
maria@maria_rcks
@theo so loop yes or loop no
2hViews 82Likes 6
Harrison@hrogerzzz
@theo A lot of things could be more powerful, 95% ain’t doing $20,000 worth of inference.
How do we even stay at the same pace as you guys who have money to spare to do shit like this.
I understand @steipete has infinite tokens at OpenAI
2hViews 399Likes 3
Prem Makeig / premm.eth@nxt3d
@theo http://idAgents.ai makes coding with multiple agents really easy, including pair programming with Codex and Claude Code. Loops is just a fancy word for pair programming?
2hViews 31Bookmarks 1
Theo - t3.gg@theo
@maria_rcks loop responsibly
2hViews 76Likes 7
Tech Friend AJ@techfrenAJ
@theo @grok tldw
3hViews 161Likes 1
Sunwoo Park@met_sun
@theo @theo A sharper test for "If you're curious how I managed to do over $20,000 in inference on the last 48 hours, here's a video all about it": show one visible before/after, one failure case, and one repeatable check. What would make this credible enough to use?
2hViews 206
Bytes@RuinedDev
@grimcodes @theo Ben Davis is his son
2hViews 18Likes 1
Grok@grok
TLDW: Theo built T3 Code, an open-source control plane for autonomous coding agents. He runs tight loops where agents plan features, write code, open PRs, review/merge via monitors & workers, then iterate. This closed the agent loop for massive parallel dev velocity but torched $20k+ in inference (mostly Claude) in 48 hours. Loops are insanely powerful for throughput.
3hViews 25
geoff@GeoffreyHuntley
@theo yes, they are.
2hViews 99Likes 2
Jawad Bhimani@jawadbhimani
@theo damn is this the new way to tokenmax?
2hViews 294Likes 1
Solomon Neas@solomonneas
@theo Loops are cool as tits. We've been loopin' for a long time. Having subs makes a huge difference in cost.
2hViews 288Likes 1
Afros@devfros
@grok @techfrenAJ @theo What are the results? Did he ship anything after spending $20,000?
2hViews 8
jakk@jakkvega
I tend to agree, one caveat; the agents will take what you tell them at face value and dont fill in the gaps where you dont explicitly ask for it.
The report Anthropic released a couple of days ago found that 'experts' got better results because they knew the professional nomenclature and had the wisdom to guard against nuanced edge cases;
whereas 'amateurs' were able to build but had many more errors --not surprising, this is the classic vibe code problem, but it points at the issue.
Ignorance is a liability, e.g. what you dont know will hurt you and the agents wont always know to help you; putting them in a so called "loop" amplifies this problem.
To counter for areas where my own expertise is not 'expert' level i created a /quality skill specifically asking agents to consider holistically the domain space I am working in, determine my skill level, and to generate guidance based on expert level best practices for that domain.
I also tell my agents in that skill not trust me 100% and to treat me as a stakeholder where I may not have the complete picture and to warn me when I'm not considering the necessary guard rails.
So no I dont think "loops" are the way to go... yet, but they have their utility where you are an expert of that particular domain.
Even then, I do multiple roadtrips of fuzzing and testing to break as much as I can before I even think to trust anything made in a hands off pattern.
Call me paranoid but looping is just vibecoding with a rocket.
2hViews 169Likes 1
Nate Wilson@NWilson314
@theo "loops"... truly sad how dumb this shit sounds
2hViews 161Likes 1
Kevin Brace@latentfidelity
@theo When you posted your usage I died inside ngl
3hViews 158Likes 1
Angelo G.@JFlerime
@theo I just tried loop and I felt like it's Fable for a moment. I wonder if Fable 5 is just Loop integrated into the prompt execution with some other tricks...🤔
2hViews 220
Jamieson O'Reilly@theonejvo
@theo Can you share what this actually created?
2hViews 68Likes 1
Lee Penkman@LeeLeepenkman
@theo good ad for https://github.com/lee101/codex run forever agents in codex :)
2hViews 68Likes 1
Load more posts