Bash Is All You Need: AI Agents Return to Unix Fundamentals
The Unix Philosophy Strikes Back
Guillermo Rauch, CEO of Vercel, crystallized what many developers are discovering through hands-on experience with AI agents:
"The primary lesson from the actually successful agents so far is the return to Unix fundamentals: file systems, shells, processes & CLIs. Don't fight the models, embrace the abstractions they're tuned for. Bash is all you need."
This insight resonates because it explains why Claude Code and similar tools feel so natural—they're not reinventing computing, they're leveraging 50+ years of battle-tested interfaces that LLMs have deeply internalized from training data.
Claude Code: Early AGI or Token Furnace?
The Claude Code discourse has reached a fascinating inflection point. Users report transformative productivity gains while simultaneously hitting practical limits:
Tobi (@tobi_bsf) captures the tension perfectly:"Using Claude Code for a week now and it genuinely feels like early AGI. The gap between 'what I can imagine' and 'what actually works' has never been smaller. But: Token consumption is insane. Running a personal assistant 24/7 hits limits fast."
He identifies the two forces that will unlock ubiquitous AI assistance: smarter token management and falling model prices—both already in motion.
The Search Problem: grep Is Dead, Long Live Semantic Search
Multiple posts highlighted that traditional code search is the bottleneck in AI-assisted development:
Rayane (@RayaneRachid_) advocates for mgrep:"grep is a tool from 1973 that does exact keyword matching. You search 'auth'? It returns EVERY line with 'auth'. Even comments, even irrelevant matches. And if your code uses 'login' or 'signin' instead... it finds nothing."
Semantic search tools like mgrep and Cass (mentioned by @doodlestein) understand intent rather than just matching strings, potentially halving token usage while dramatically improving search accuracy.
Power User Techniques Emerge
Practical tips for Claude Code optimization are spreading:
- Jarrod Watts shares his
comments.mdfile that prevents Claude from writing "slop comments like 'increment counter' on already self-documenting code" - Andrew Jiang published the "Idea" skill—tell Claude an idea, it spins up tmux, researches it autonomously, then sends results to Telegram
- 0xSero recommends git worktrees for running parallel agents on long tasks
- dan (@irl_danB) promotes OpenProse as "the most powerful agent orchestration pattern"
Context as Capability Unlock
Danielle Fong articulates something profound about AI assistance:"If you get all your notes into the AI it dramatically enhances its own ability to 'get' what you want because it has associative memory and access to that. It is a massive capabilities unlock... for the things I have been dreaming about I can reference them by gesturing at it."
The implication: personal knowledge bases aren't just nice-to-have, they fundamentally expand what AI can do for you.
Small Models, Big Impact
@Hesamation highlights research from Amazon and NVIDIA showing small language models can outperform 500x larger models on agentic tool calling with proper fine-tuning:"Agent-focused companies must adopt more development of LLMs this year. They have the data and the right playgrounds. It's economically senseless to use proprietary large models in most agentic use cases."
From Demo to Production
Nina (@HeyNina101) reminds builders that weekend demos need four distinct architectural layers to become production applications—a sobering but necessary perspective as the ecosystem matures.Resources and Learning
Pamela Fox updated her guide on keeping up with gen AI news, recommending bloggers like Simon Willison, Gergely Orosz, Hamel Husain, and Gary Marcus. Neo Kim compiled 12 essential system design case studies covering Google Docs, Spotify, Reddit, Kafka, and more—increasingly relevant as AI applications need to scale.---
The throughline today: AI agents work best when they embrace existing computing primitives rather than fighting them. The winners will be those who master the art of providing context, managing tokens, and letting models do what they're already great at—working with files, shells, and text.Source Posts
M
If you feel like you're caring less about code quality in the AI age, read this
M
hyperliquid can now scale to trillions
because people can finally vibe code apps on it
the new financial system is now open to the public with easy apis any ai can understand
hyperliquid https://t.co/oRNiFMcSwe
ℏ
that's it... Amazon proved small language models can outperform 500x bigger LLMs in agentic tool calling. NVIDIA already showed this idea before, and this is another proof that the right fine-tuning can make SLMs better in agentic applications than LLMs, with a fraction of the cost.
agent-focused companies must adopt more the development (dataset, evals, post-training, design) of LLMs this year. they have the data and the right playgrounds for these LLMs. it's economically senseless to use proprietary large models in most agentic use cases.
read this paper and also NVIDIA's ("Small Language Models are the Future of Agentic AI") to be convinced.
A
Published the Idea skill!
v1.0
- Tell @clawdbot "idea: [brilliant idea]"
- Spins up tmux and a Claude Code session
- When the research is done, the research file gets saved locally and sent to your Telegram
Uses tmux and telegram cli tools
https://t.co/a2JkmiAfkQ
C
The woman in Minneapolis died because she thought she could play chicken with federal law enforcement due to being completely divorced from reality.
Life isn’t GTA, this isn’t a movie, everyone is trying to determine if she meant to hit the agent or not when the question really is… Why did she feel emboldened enough to blockade on-duty federal law enforcement and flee when explicit orders were given to stop?
B
This got me interested in Tinker and so we made a Claude Code skill for it
https://t.co/lO8GnmsOSC
d
if you haven't tried OpenProse yet you really should
you are sleeping on the most powerful agent orchestration pattern and the only of one of its kind
/plugin marketplace add git@github.com:openprose/prose.git
/plugin install open-prose@prose
-- restart Claude Code --
"use prose to automate X"
N
If you're serious about SYSTEM DESIGN (in 2026), learn these 12 case studies:
1. How Google Docs Works
↳ https://t.co/W57IkAjXpT
2. How Spotify Works
↳ https://t.co/BxrH3oHIFS
3. How Reddit Works
↳ https://t.co/o6Pw2hhj3T
4. How Bluesky Works
↳ https://t.co/2rLYlRlky0
5. How ChatGPT Works
↳ https://t.co/5lCKxq2g4N
6. How Kafka Works
↳ https://t.co/8rOy9KgCMo
7. How Slack Works
↳ https://t.co/eIo29uOQOJ
8. How Meta Handles 11.5M Serverless Function Calls per Second:
↳ https://t.co/NSt6jovxu5
9. How Uber Finds Nearby Drivers
↳ https://t.co/kJ2t8dtmch
10. How Twitter Timeline Works
↳ https://t.co/pF2RYmPaIG
11. How YouTube Was Able to Support 2.49 Billion Users With MySQL
↳ https://t.co/4VDJ5cs6fL
12 How to Scale an App to 10 Million Users on AWS
↳ https://t.co/RozCGli0r8
(What else should make this list?)
——
👋 PS - Want my System Design Playbook for free?
Join my newsletter with 200K+ software engineers:
→ https://t.co/ByOFTtOihX
———
💾 Save this for later, and RT it to help others master system design.
👤 Follow @systemdesignone + turn on notifications.
E
Two unpopular opinions on the Minneapolis shooting:
1. The woman who died was at fault. She was acting erratically, inserted herself into a dangerous situation, and behaved in a way that put both herself and others at serious risk.
2. The ICE agent acted on instinct and fired too quickly. He should not have escalated to lethal force in that moment. We need full body cam footage to understand exactly what he perceived, but based on the video available, it does not appear his life was in immediate danger as the vehicle was moving away.
The left is already calling this blatant murder.
The right is already claiming the woman was trying to kill the agent.
Neither narrative is accurate.
The woman who died acted dangerously and recklessly, and the ICE agent drew and fired his weapon too fast.
If both sides don’t pause the political posturing and focus on de-escalation, Minneapolis is headed toward another night of chaos just like George Floyd.
ICE should temporarily pull back from the community, and leaders on the left need to clearly and forcefully call for calm and nonviolence.
There’s almost no chance both of those things happen, which means this situation will likely get ugly and Minneapolis could burn again.
I know this take won’t make me popular with either side but it’s just my initial thoughts on situation.
S
This is literally how Boris, the Creator of Claude Code uses Claude code
Bookmark for later https://t.co/7Eq1OYOe7h
G
The primary lesson from the actually successful agents so far is the return to Unix fundamentals: file systems, shells, processes & CLIs.
Don't fight the models, embrace the abstractions they're tuned for. Bash is all you need.
P
I always get asked how I keep up with gen AI news, so I've updated my blog post about "how I learn gen AI" with my fav bloggers.
https://t.co/3oElwllkr7
You can also follow them here: @simonw, @GergelyOrosz, @HamelHusain, @intellectronica, @isaac_flath, @GaryMarcus https://t.co/02etDiN3zT
N
To move from a weekend AI demo to an AI production-grade application, you need to architect these 4 layers 👁️👁️ https://t.co/EzQDEqt8kZ
D
listen, if you get all your notes into the AI it dramatically enhances its own ability to "get" what you want because it has associative memory and access to that. it is a massive capabilities unlock to have this. for the things i have been dreaming about i can reference them by gesturing at it.
and the problem of handling a mountain of context is inherent to the process, because it eventually creates it. it's perfectly reasonable people are bootstrapping and dogfooding here
M
how to actually use ai and claude code to automate your trading
cars drive themselves, why are we still trading by hand?
if you want to remove emotions and get your time back
this is a free training showing you step by step how to build bots
1 of my last ever btw
war time https://t.co/TtKtGK833j
T
Using @clawdbot for a week now and it genuinely feels like early AGI.
The gap between "what I can imagine" and "what actually works" has never been smaller.
But: Token consumption is insane. Running a personal assistant 24/7 hits limits fast.
Two things need to happen:
• Agents get smarter about token usage (just shipped a PR reducing prompt overhead)
• Model prices keep dropping (already happening - GPT/Claude/Gemini racing down)
The future where everyone has a personal AI running in the background isn't limited by capability anymore. It's limited by cost.
That window is closing fast. 👀
r
How to stop feeling behind with AI (practical guide)
R
Bon parmi les tools que j'utilise sur mon setup Claude Code et qui pour moi est INDISPENSABLE, c'est mgrep !
Mgrep c'est quoi ?
- De base, Claude Code utilise grep pour chercher dans ton code. Grep c'est un outil de 1973 qui fait de la recherche par mots-clés exacts. Tu cherches "auth" ? Il te sort TOUTES les lignes avec "auth". Même les commentaires, même les trucs pas pertinents. Et si ton code utilise "login" ou "signin" à la place... bah il trouve rien
Le problème c'est que Claude passe son temps à scanner des centaines de fichiers pour rien, il brûle des tokens, et souvent il rate ce qu'il cherche vraiment
Mgrep c'est bien plus puissant. Tu décris ce que tu veux en langage naturel : mgrep "où est-ce qu'on gère l'authentification ?" et il comprend le SENS de ta recherche. Il te sort les bons fichiers direct, peu importe comment le code est nommé
Résultat : ~2x moins de tokens, des recherches way plus précises, et Claude qui passe son temps à réfléchir au lieu de fouiller dans le vide
J
A simple but very effective rule I use in Claude Code is this "comments[.]md" file.
It prevents Claude from writing slop comments like "increment counter" on already self-documenting code. https://t.co/WYDG7HPcCu
J
@steipete @raw_works @dhh @clawdbot Cass is much faster because it indexes the stuff ahead of time. And it’s a proper search engine with ranking, lexical/semantic search, etc. And supports every coding agent out of the box. You should give it a try, Claude is very good at using cass in its robot cli mode.
L
Everyone who tried Claude Code over the holidays and has been thinking deeply about this for 7 days:
All white collar jobs are toast. It is over
0
There you go, open source.
If you want to get into running parallel agents and experimenting with long running tasks, I recommend git worktrees.
This is built into cursor but now you can get it in Opencode!
Your model can spin up worktrees and assign subagents to work on them (:
https://t.co/be7OLqXLnD