
18 y/o developer, built a thing which got 15k+ users, contributor @t3dotcodes
Joined January 2026
- Tweets1.3K
- Following147
- Followers143
- Likes3.5K
Utkarsh retweeted
Replying to @jullerino @t3dotcodes
Thank you @UtkarshUsername for doing the grunt work for this feature!
Today we’re releasing Sarvam Prism, a System One model built on Sarvam 105B for fast, accurate decisions across enterprise workflows.
It scores 71.8 with a median response time of 44 milliseconds, the highest average score among decision models that respond in under 100 milliseconds.
Utkarsh retweeted
Made a thing: @t3dotcodes widget for Noctalia. I find it handy, maybe you will too :)
github.com/piexelated/agent-…
Utkarsh retweeted
At this point the reaction is:
Bad software: I'll just build my own.
Good software: I'd rather use this and let it get even better than build my own.
Utkarsh retweeted
So apparently @t3dotcodes was the fifth most engaged open source project this year. Pretty cool 🙌
another thing that came to mind right now, is anyone using any TUI DB manager such as LazySQL, Rainfrog, etc..
if you do, I would appreciate feedback on what you like, don't like, think is missing, think is bloat, etc..
i'm making an actual good TUI database manager. i don't think any of the ones i tried either work as well or have all the features i want. still very early in dev, but if you wanna follow along; github.com/0x4bs3nt/pond
Utkarsh retweeted
T3 Code originally was such a light wrapper that your agent didn't even know what it was.
With Orchestrator V2 (harnesses can call each other), we had to fully overhaul this.
With built-in browser, device manager for mobile, subagents/delegates, "send to other thread", html inline visualizations, etc...it got a bit bloated.
Biggest issue first: we listed all the models available to use for subagents. For most people, didn't matter. For OpenCode and Pi users with literally hundreds of models? Real bad. Fixed now, sorry!
I knew I had to go further. The tool schema was...too big. Far too big. So I took that as a challenge.
I created an insane eval. Spent like a week on it. Took all the ways my agents used these tools in my day to day work, as well as all the failures and rage moments I had, and turned them into a generic eval I can run to test changes to the tools and their descriptions.
I expected to mostly focus on trimming the descriptions, but I was AMAZED how many error cases I found. My eval had over 80 cases, and Opus failed over half of them.
As you can guess, the hill climb began soon after :)
Step 1: Bench Sonnet 5.5
Step 2: Keep adjusting tool descriptions until Sonnet 5.5 gets perfect score
Step 3: Add Sol into tests, do the same
Step 4: Add Opus, do the same
Step 5: merge it (because error rates are better now)
Step 6: new thread, tell Opus to trim until it uses 50% less context without any regressions in the eval
Step 7: ship it
This was super fun. Amazed at how many problems you can kinda just turn into evals and then do shit like this. The "shape" of this was not far from the UI overhaul hill climb. Surreal you can just, like, do this.
Utkarsh retweeted
Fwiw, I think the "realistic pareto line" right now is:
- GPT-6 Luna
- GPT-6.1 Sol
- Opus 5.5
Utkarsh retweeted
T3 Code Nightly users who just updated: I'm sorry. We shipped a broken one.
Revert to stable for now. Stay there if you want fewer breaks (probably for the best). New build will be live in 10 or so mins to fix this.
Utkarsh retweeted
We've started adding some compaction to the t3 database, but old databases does not release the freed up space to the OS.
You can run `t3 doctor` to rebuild the database which will allow future compactions to release freed up space to the OS automatically. Also suggests you to clean up the older v1 database when you've made sure v2 works for you and you don't need to roll back.
Utkarsh retweeted
Replying to @harshilmathur
The amount of time and effort it takes for a person to bake is very different from what's required to do something with AI.
Utkarsh retweeted
T3 Code reached 400k users last night, that's cause for some celebration don't you think?
Utkarsh retweeted
more @t3dotcodes changes hit nightly:
⚡ cmd/ctrl+enter queues threads in the background by @ex0t1clol
🧩 skills now show up in the / menu (me)
🖼️ choose project icons from anywhere on desktop @bil0090
💎 cursor now appears when installed (rishet11 on gh)
✨ easier theme search, thread renaming @UtkarshUsername + a ton of qol fixes
more @t3dotcodes changes hit nightly:
🛠️ tool calls collapse into one line with live status
🧩 approvals, questions, plans, tasks & stashes
🧭 refreshed workspace nav, usage + pr uis
✨ cleaner thread menus + terminal close confirmations
➕ a ton of ui fixes
update if you haven't
Utkarsh retweeted
TRUST MAXXXX as much as you can.
Either you like it or not. you need to build trust with people.
can one trust you to do something ?
like ik developers from the t3code community, i will anytime hire them if i need any dev in the future!
@UtkarshUsername @pcstyle53
@gilrdb for the design work.
these are some of the examples. ik them, i have yapped with them. ik they are good at their shit. they have built the trust with me now.
and this shit doesn't happen by faking it, it happens with genuine curiosity. like they don't talk to me thinking, oh i am gonna talk to him this way, so in future he hires me or like helps me. we genuinely get excited over shit and talk about shit.
PLS DON'T TRY TO LARP. you can spot a larper as you can spot a genuine person.
oh wrote a lot. but true tho.
Replying to @RhysSullivan
anyone talking about their custom setups to make ai more effective isn't doing anything useful
coding used to be manual labor. typing the code out, doing large changes, jumping between files. a fancy, bespoke setup could make a difference
but it's not like that anymore, you have a magic box you can say stuff into and things happen
people haven't understood this, they're still spending their energy tinkering with what's behind the box. the underlying tone is always frustration. they're clearly not happy and maybe even on a burn out path
the best people at this are doing the best work of their life all while happier than ever they could do this even on a vanilla setup
the box can already do amazing things. it can even produce the best code of your life. it's limited entirely by the person using it and their creativity and their ambition
and you're not gonna get that in a skill file




















