Pinned Tweet
Everyone wants to do Code Mode agents
No one is truly embracing Code Mode
So I made a new agent with Code Mode as the foundation
Meet OC++ (a fork of @opencode)
youtu.be/RTJZRF2sotg?si…
No one is truly embracing Code Mode
So I made a new agent with Code Mode as the foundation
Meet OC++ (a fork of @opencode)
youtu.be/RTJZRF2sotg?si…
7
3
65
35,726
You will pry away subagents from my cold dead hands
Anthropic is now recommending that companies REMOVE unnecessary subagents from their AI systems.
This is a pretty significant shift in agent architecture, with their team finding that a single agent equipped with skills can frequently outperform complex multi-agent systems, and one customer seeing roughly 2x faster execution.
The emerging rules of thumb:
→ Skills: Teach the agent HOW to do something, including procedures, expertise, and instructions.
→ Tools / MCP: Give the agent the ability to DO things through APIs, actions, and live data.
→ Subagents: Use for parallel work, isolated context, or tasks a single agent struggles with.
The interesting part is progressive disclosure. Instead of spawning specialized agents or stuffing everything into a giant prompt, one agent discovers and loads skills and tools as needed.
Their advice is to fold existing subagents into the main loop, turn their instructions into skills, expose their actions as tools, and run evals to see if performance holds up.
As models have gotten smarter, improving agent architecture is becoming less about what you can add and more about what you can delete.
claude.com/resources/arti…
0
0
0
395
Now think about the equivalent for front end frameworks.
It has to be fast, small design space, and structurally verifiable
Hmm I wonder where you can find that
It has to be fast, small design space, and structurally verifiable
Hmm I wonder where you can find that
The next great "language for agents" should be:
- As fast as Rust
- As little design space as Go
- Have the strictest type/verification system ever written
Bend seems like an interesting first effort
2
0
25
5,509
First time I’m hearing about someone feeling guilty about being a guy in tech
Not the first time I’m seeing someone do literally nothing to diversify tech
Not the first time I’m seeing someone do literally nothing to diversify tech
"Yes, men dominate computing. There are many reasons for that; none of them justifies collective guilt. Turning the percentages into a struggle session is nonsense." world.hey.com/dhh/be-a-tech-…
10
2
93
8,990
The data loading and caching story of both ChatGPT and Claude are horrendous
I was really expecting better from trillion dollar companies but when you let AI drive the implementation without proper investment into infrastructure I guess this is doomed to happen
Please use a proper sync engine folks
I was really expecting better from trillion dollar companies but when you let AI drive the implementation without proper investment into infrastructure I guess this is doomed to happen
Please use a proper sync engine folks
0
0
9
826
Maybe instead of emphasizing the skill level of the hacker you should highlight how bad the security of the university infra should be to allow this
Didn’t even require Opus or Sol to break in
Didn’t even require Opus or Sol to break in
‼️ A low-skill hacker used a jailbroken DeepSeek V4 to breach an Indian public university and steal 5,200+ student and faculty records, Unit 42 says.
The AI agent even drafted a "NEXT EXPLOIT OPPORTUNITIES" plan.
Unrelated criminals now run the same cheap DeepSeek setup.
0
0
19
1,292
Dev Agrawal retweeted
MiniMax-M3.1-Flash is open source! 🚀🐰
428B parameters. 1M context window. Native video input support.
• BankerToolBench: 76.2
• DeepSWE v1.1: 73.7%
• OSWorld V2: 61.7
• Agents’ Last Exam: 29.6
• CyberGym: 82.2
Hugging Face: huggingface.co/MiniMaxAI/Mini…
428B parameters. 1M context window. Native video input support.
• BankerToolBench: 76.2
• DeepSWE v1.1: 73.7%
• OSWorld V2: 61.7
• Agents’ Last Exam: 29.6
• CyberGym: 82.2
Hugging Face: huggingface.co/MiniMaxAI/Mini…
0
1
4
268
Dev Agrawal retweeted
TanStack AI 1.0 is officially stable! 🎉
Build chat, voice, image generation, agents, and harnesses into your apps, with your choice of models, UI framework, and infrastructure.
Your AI, your way.
tanstack.com/ai
Build chat, voice, image generation, agents, and harnesses into your apps, with your choice of models, UI framework, and infrastructure.
Your AI, your way.
tanstack.com/ai
83
121
1,487
93,661
There is a massive difference between agents writing 99% of your code and 100% of your code
That 1% of code you write yourself can do wonders to improve the other 99% provided you know what you’re doing
That 1% of code you write yourself can do wonders to improve the other 99% provided you know what you’re doing
4
3
38
2,131
Turns out we shipped slop code way before LLMs existed
we’re about to discover that 20 years of software engineering was largely an exercise in accidental complexity, architectural overfitting, and institutionalized mediocrity. most developers optimized for shipping code, not understanding computation.
2
1
12
1,440
you want liveview in js, but actually good? we got it!
The stream is back tomorrow. I'm going to talk about my exploration into the last quadrant of my unified architecture grid: Live Server Components.
youtube.com/live/Sr1z-u3Ux…
1
0
5
562
Dev Agrawal retweeted
The stream is back tomorrow. I'm going to talk about my exploration into the last quadrant of my unified architecture grid: Live Server Components.
youtube.com/live/Sr1z-u3Ux…
youtube.com/live/Sr1z-u3Ux…
5
4
46
4,512
live now! hop in and learn why @solid_js is the framework your agents should be using
Reminder if you want to learn about solidjs. Ill be speaking here on Oct 22nd
luma.com/techtank-frame…
0
0
5
492
Dev Agrawal retweeted
Reminder if you want to learn about solidjs. Ill be speaking here on Oct 22nd
luma.com/techtank-frame…
luma.com/techtank-frame…
0
4
19
1,108
powersync continues to be the most sensible sync solution and it's not close
let’s talk about sync
ever since watching theo’s videos about t3 chat local sync, I wanted it for smry.ai
but as I worked on it, I realized it’s not a good-to-have, it’s a must-have for a reading assistant, as your reading history needs to load instantly.
also users want to own their data, which they do if it’s cached on their own browser
i evaluated every provider: zero/electric/etc, even built my own shitty one!
but nothing was as good as powersync, which is super cheap, scales like crazy, has a drizzle integration, and has a team with actual support (also everyone else in the space seems to die/get bought)
powersync was the best solution for sync, but really the team was the best part
i’m able to work much faster now that I’m not worried about hand-rolled sync breaking with every change, and with all code written by agents it’s good to have a tool that has the things they need to succeed
0
0
3
578
Dev Agrawal retweeted
I always come back to this writeup from @Rich_Harris. It's been the oracle of most JS framework compiler work the past decade. This is why others failed where Svelte succeeded.
AI suggests we rethink this. We want restriction, not extension.
gist.github.com/Rich-Harris/0f…
AI suggests we rethink this. We want restriction, not extension.
gist.github.com/Rich-Harris/0f…
11
11
159
10,951
