Andrej Vasiljev

[Dean W. Perkins @deanwperkins]

Anthropic just released a 37-minute guide on how to build AI agents that can automate an entire business.

It’s free, and it comes straight from the engineers who built Claude.

Agents that actually work, delegate tasks, and get things done on their own.

Save this post 🔖

[marcus @marcusyul]

ALGUIEN LLEVA MESES DOCUMENTANDO CÓMO USA CLAUDE CODE EL EQUIPO QUE LO CONSTRUYÓ, Y LO TIENE TODO EN UN REPO GRATIS

83 tips recogidos de Boris Cherny, Thariq y el resto de Anthropic.

No teoría. Cada concepto viene con su implementación funcionando dentro del repo.

→ Subagents, commands, skills, hooks y MCP con ejemplos reales
→ 12 workflows comparados: Superpowers, Spec Kit, BMAD y más
→ Agent teams con tmux y git worktrees para ir en paralelo
→ Cross-model: planificas en Claude Code, revisas en Codex

Tres tips que solos ya valen el clone:

→ CLAUDE md por debajo de 200 líneas.
→ Subagent con contexto fresco antes que compactar.
→ El context rot entra sobre los 300k tokens, no dejes que la sesión pase de ahí.

Si usas Claude Code como un ChatGPT con terminal, te dejas media herramienta sin tocar.

65.415 stars y licencia MIT.

🔖 Guárdalo, te servirá. Link abajo.

[elvis @omarsar0]

This WikiSkill paper from Google is a must-read.

At a high level, it shows the effectiveness of persistent agents, knowledge bases, and skills.

@karpathy popularized LLM Wikis.

But this paper provides an actual framework for how agents can tap into a wiki of skills that evolve.

What's fascinating to me is how this can complement your agents.

LLMs can only learn so much about the world. External knowledge is crucial to get agents to do tasks efficiently and accurately in the real world.

So this is why I think this paper is an important one, as it tries to fix some of the common issues you face when building and maintaining skills.

It automatically leverages your agent runs, persists that knowledge into a wiki, and uses all of that to keep skills properly tuned for reusability.

The most impressive part of WikiSkill is that it appears to be model-agnostic. In other words, it works across different tasks and models.

The evolved skills can even transfer to smaller models that sometimes outperform bigger models. This hints at the effectiveness of persistent agents, via persistent knowledge bases and evolved skills.

The big question for me is how evolved skills coming out of WikiSkill transfer to the next generation of models. I think they will provide a huge advantage and be leveraged in more interesting ways by smarter models.

The practical takeaway here is that we should all be thinking about how to build persistent knowledge bases across our companies and projects. And how to use that to upgrade and evolve our skills.

Join our community to discuss this paper more: https://t.co/AsWbH4jzgn

[Bober_smart @Bober_smart]

I still genuinely wonder why not everyone is using this approach yet. After all, it was thanks to this exact scheme that I managed to reach a $12,000 income six months ago

Elon Musk published a simple idea that hit 30 million views: the most likely outcome is that AI and robots make everyone wealthy. In fact, far wealthier than the richest person on Earth

This is precisely what Grok Build is, and the smartest configuration currently costs $0 extra:

Create a bot, name it Brain

Connect it to a local folder and tell it to build a private wiki of everything you know right on your computer

It will create key folders: raw for sources, wiki for its pages, and save the entire method as a working skill

Drop anything into the chat: an article, a PDF, a YouTube link, or a screenshot it will instantly read, link, and write it down

Set up routine ingestion: while you sleep, the system pulls in everything you saved to your bookmarks

Ask questions about absolutely anything you have ever saved, forever

Your laptop stays closed, but your knowledge base grows on its own. Just five minutes to set up - and this process starts compounding like interest from day one

Not using Grok in 2026 is a huge mistake, just as big of a mistake as not bookmarking this

[elune @elune0x]

Andrej Karpathy spent 8 years at OpenAI and Tesla

Last week, he compressed everything he knows into one free 2-hour lecture

Agents → Harness → Loops → Graphs → Self-Improving Systems

People pay $15K for bootcamps that teach less than this

This lecture is better than most paid AI engineering courses

You probably don't have 2 hours right now

Don't let this disappear from your feed

Watch it

Then read the guide on harness engineering below

[arc. @arceyul]

GitHub acaba de solucionar el mayor problema del vibe coding.

Acaban de lanzar Spec Kit y en días ya tiene +95K estrellas.

¿La idea?

En vez de tirar prompts vagos y rezar para que el agente no rompa tu proyecto…

Spec Kit obliga a la IA a crear una especificación estructurada ANTES de tocar código.

La IA primero entiende lo que quieres construir, pregunta lo que falta, organiza el proyecto y después empieza a programar.

Eso significa menos tiempo arreglando errores absurdos, menos código inconsistente y resultados mucho más predecibles cuando trabajas con agentes.

El flujo es simple:

/constitution → reglas y estándares
/specify → qué quieres construir
/clarify → dudas antes de empezar
/plan → arquitectura y stack
/tasks → tareas ordenadas
/implement → ejecución

Compatible con Claude Code, Cursor, Copilot, Codex, Gemini CLI y +25 agentes.

95K estrellas.
8K forks.
Open source.
Publicado por GitHub.

REPO DEBAJO ⬇️

[lumxss @bkdgiffug]

剑桥这回直接扔王炸了！！

AI & ML经典教材全集直接免费开放，PDF随便下。

想学机器学习又不想被割韭菜买高价课的，这十本刷完，底子基本就硬了。

顺序从易到难排好了：

1️⃣ 《机器学习理解》——理论算法一把抓，零基础入门首选
🔗 https://t.co/fylTw37bOl
2️⃣ 《机器学习数学基础》——数学底子弱的先把这本补上
🔗 https://t.co/yykNLQmdfv

3️⃣ 《机器学习算法的数学分析》——深入数学原理
🔗 https://t.co/MEvgkXoHYy
4️⃣ 《深度学习理论原理》——搞懂DL背后的理论根基
🔗 https://t.co/ig93KOQRPn

5️⃣ 《神经网络与机器学习》——神经网络的系统讲解
🔗 https://t.co/cTMBvJz6Ny
6️⃣ 《图深度学习》——图神经网络入门必读
🔗 https://t.co/RJeCqCstml

7️⃣ 《机器学习的算法视角》——从算法角度重新理解ML
🔗 https://t.co/VT0YnKdUdO
8️⃣ ��概率论：理论与实例》——概率基础打牢
🔗 https://t.co/NqVoIZGGYg

9️⃣ 《应用概率基础》——概率论实战应用
🔗 https://t.co/FrxqQER3mQ
🔟 《高级数据分析》——数据科学进阶必备
🔗 https://t.co/Zt74B21nrf

说句实话，这些书没一本是轻松的，别指望躺着翻完。
但只要你能硬啃下来两三本，比听群里吹一年AI牛逼都管用。

[shiqway92 @shiqway92]

THESE 3 GITHUB REPOSITORIES GIVE AI AGENTS A MEMORY THAT SURVIVES BETWEEN SESSIONS

Together, they have around 129K GitHub stars.

1. Claude-Mem - 92.6K⭐

Captures what an agent does, compresses its history, and restores relevant context in future sessions.

https://t.co/g8P7czZOFp

2. TencentDB Agent Memory - 25.2K⭐

Creates shared memory for teams of agents using conversations, documents, skills, and codebase knowledge.
https://t.co/gnZoXLJnaP

3. Semantica - 11.3K⭐

Turns an agent’s knowledge, context, and decisions into a traceable knowledge graph.
https://t.co/t6DbrWPsCW

Together, these projects gained approximately 22K new stars in august.

Ai agents become much more useful when they stop starting every session from zero

[Andrej Karpathy @karpathy]

One pattern I find useful for working with LLMs is a nice long ramble session. Sometimes the LLM needs more bits to understand what you're trying to achieve, but you're too lazy to type them. In these cases I like to lean back, switch to /voice and just ramble for like 10 minutes, total mess, anything goes, full stream of consciousness. Sometimes I declare it up top, something like "switching to speech recognition sorry for any typos...". Sometimes I turn it into a small interview of a few turns. But I find that the LLMs are somehow very good at reconstructing long incoherent rambles and often their echo of your own tangle of thoughts comes out quite a bit cleaner than what you started with. The result is that you improve the mind meld and have to correct things less from that point on.

[rvaniaaa @rvaniaaaa]

a researcher showed me this architecture and said one thing

“i stopped building two separate systems the moment i understood this”

most people building AI systems make one of two mistakes they build a second brain that remembers everything and acts on nothing or an agent team that executes fast and forgets everything by morning one stores knowledge the other produces action neither compounds

this architecture connects them

six agents one brain one loop

> SCOUT finds signals and passes them forward
> ANALYST builds context from everything the brain already compiled
> STRATEGIST designs the approach from memory not from scratch
> EXECUTOR acts within approved boundaries
> GUARDIAN enforces the rules before damage happens
> OBSERVER closes the loop and feeds every outcome back into the brain

the part most people miss is what happens between cycles inputs go in raw the brain compiles them agents read compiled knowledge not raw data outcomes feed back into the brain the next cycle starts with more context than the last

“all agents read from the second brain before every move nothing starts from zero”

that one line is the entire point

the gap between a system that executes and a system that compounds is OBSERVER most people build five agents and skip the sixth they get automation they never get compounding

the sixth agent is the difference between a tool and a system that thinks

full article below every agent prompt included

[kocer @kocer_eth]

FIVE LAYERS OF AGENT ENGINEERING, EACH ONE WRAPS THE ONE BELOW IT. IF YOU SKIP LAYER 2, YOUR LAYER 5 WILL LOOK BROKEN WHEN IT IS ACTUALLY JUST STANDING ON NOTHING.

for weeks i debated harness vs loop vs graph like they were competing choices.
then a stack diagram made the shape obvious. they are not choices. they are floors.

01 | prompt engineering. the message.
unit of work: one input. inputs are role, instructions, examples, format. output is a single raw response.

02 | context engineering. the memory.
unit of work: what stays in the window. a curator selects, compresses, and drops from query, docs, memory, prior turns, and tool outputs before the prompt runs.

03 | harness engineering. the machine.
unit of work: the machine itself. gather (context + prompt) → LLM → tools or sub-agents → verifier → final response. the article calls this the operating environment.

04 | loop engineering. the system.
unit of work: the run. goal + success criteria + max iterations + budget + completion check wrap around one harness pass. failed pass appends results to context and retries.

05 | graph engineering. the topology.
unit of work: the graph run. goal + nodes + edges + state schema. graph routes to agent nodes, tool nodes, or human approval. a reviewer node with a different model and fresh context checks the final answer.

the wrapping is the whole point. layer 5 assumes layer 4 works. layer 4 assumes layer 3 works.
skip layer 2 and layer 3's verifier keeps failing without a clear reason.

this is why swapping the model is a one-day project and swapping the stack is a quarter.
the model is the commodity. the five layers around it are the engineering.

full three-layer breakdown of the top of the stack (harness, loop, graph) in the post below.

[Yarchi @undefinedKi]

The winner of an Anthropic hackathon open sourced his entire Claude Code setup. 68 subagents, 286 skills, 94 commands, MIT license

ECC turns Claude Code from one assistant into a full engineering team. It plans before it builds, writes the failing test first, then reviews its own work from a fresh context.

The agents, by job:

Planning. Turns a one-line request into a blueprint you approve before any code exists.

Review. Reads your diff in a clean context, with separate reviewers for Go, Python, TypeScript, Rust, Java and more.

Build repair. One resolver per toolchain, down to PyTorch and CUDA errors.

Security. An OWASP pass on your code, plus a scanner that audits your own agent config for injection risks.

Architecture. System design calls before they turn into migrations.

Domain work. Database queries, ML pipelines, end-to-end tests, docs.

The skills, by category:

Testing. tdd-workflow gates you from red to green, with eval-harness and verification-loop sitting on top of it.

Language packs. Idioms, testing and security for Python, Go, Rust, C++, Django, Laravel, Spring Boot, Next.js.

Context. search-first makes it read the docs before writing, iterative-retrieval keeps subagents from dragging your whole repo into the window.

Shipping. Docker, CI/CD, health checks, rollbacks, and migration patterns for Prisma, Drizzle and Django.

Work that is not code. Writing in your voice, market research with sources, pitch decks, slide decks.

Start with one plan and one rules pack. Installing all 286 skills at once is the fastest way to make it worse.

[Mr. Buzzoni @polydao]

CLAUDE + OBSIDIAN + LOOP ENGINEERING = A VAULT THAT RUNS ITSELF

the core idea: the vault is the loop's state, not the chat window

everything Claude knows lives in a .md file

the loop:
> capture - a thought lands in 00-inbox
> context - Сlaude Opus 5 pulls links, tags, and neighbouring notes
> draft - edits happen inside a git worktree, never the live vault
> review - a critic agent checks the diff before anything ships
> commit - appended to the vault, nothing gets rewritten

the key insight: frontmatter fields like supports, contradicts, and supersedes are graph edges, not metadata - the note format is the write API

start with a plain loop, it runs about 2-4x the cost of one direct call

> only move to a full graph once state has to outlive the session, several agents need to coordinate, or you have to explain what changed - that jump can run 10-50x

one review assistant climbed from 55% to 72% to 84% just by moving through these shapes in order

worth stealing even without a graph: the review step. most vaults skip it - claude writes straight to live notes

errors compound silently for months

set that gate first

[NO1ennn @N01ennn]

I gave my Grok Bot two options: multiply or die. It came back asking for $250 to hire a team of six.

The same headcount on a real desk is three analysts and a risk officer. That is $400,000 a year before anyone opens a position. Mine cost $250 once and about $300 a month to keep breathing.

Mine took $250 to $611 on day one, out of a wallet anyone can open and read.

Most of the build is not about making it trade. It is about stopping it.

Every agent's job description ends with a list of what it may never do without asking me first.

One of the six is not allowed to trade at all. Its entire job is killing what the other five found, and its no beats every score on the desk.

The one that makes the final call never touches the market either. It only signs.

Day three the account was back to $180 and it opened nothing for nine hours while every account I follow posted green screenshots. That was the day the build started working.

You could have this running by Friday. Most people reading this will not, and in a few months they will be paying somebody who did.

The whole build is in the post below ↓

Save it while it still sounds strange. Give it a quarter and this is just how a retail account works.

[Chesny @chesny]

> empezó con $48

> ganancia total $780,000

La estrategia es simple:

→ Apunta constantemente a mercados donde aún no se han ajustado del todo

→ Entra antes de que el repricing se complete

→ Repite el mismo setup una y otra vez a lo largo de un número enorme de entradas

[Ricker @0xRicker]

Andrej Karpathy just rewrote the rules of using LLMs:

"Prompting is going away. Delete everything, keep Graph."

LLMs → Prompts → Agents → Graphs

He dropped his full 2-hour course from Stanford on "Graph-Native Research"

• 00:00 - Intro to Graph systems
• 01:08:09 - LLMs architecture

This Karpathy course can replace a $100K Yale LLM senior degree.

Watch it today, then save the full graph engineering guide below

[Norvex @norvex1029]

I LAUNCHED 400 AGENTS AND HOOKED THEM UP TO A SECOND BRAIN — AN ACTIVE CLOUD SUPERCORTEX COORDINATING THE ENTIRE DIGITAL SWARM IN REAL TIME.

Before that, though, the Second Brain did something else.

In just a couple of minutes, it tore through gigabytes of old junk that had been gathering dust across folders for years, sorting abandoned documentation, forgotten notes and archive files into place.

Then the dashboard exploded with a stream of activity.

Four hundred autonomous threads started working at once — writing code, hunting for bugs and testing each other's output.

If one agent hit a wall on complex logic, the Second Brain pulled a successful pattern from a neighboring thread and fed it back into the swarm.

A digital storm of pure parallel execution.

Four minutes flat later, the logs went quiet.

The terminal lit up:

400 tasks completed.

Zero errors.

I opened the finished project.

Clean architecture.

Complete test coverage.

A deploy ready to go.

I closed the laptop, smiled and went to sleep while the Second Brain kept working quietly in the cloud, finishing the documentation for the next sprint.

I highly recommend bookmarking this article.

[hitesh.eth @hmalviya9]

@0xparag @virtuals_io https://t.co/DjL45l9koC

[OSF @osf_rekt]

Hot take - same crowd that was in ICOs in 2017, were in DeFi in 2020, were in NFTs in 2021, were in memecoins in 2024 and are now in AI coins in 2025.

I call this the “casino crowd”. I think the vast majority of it doesn’t care/understand about the tech or culture, but really just participates to attempt to money (otherwise known as hyper gambling).

It makes some people disgustingly rich, and as such those people have the liquidity to roll the dice again…because…I guess that’s just human nature 99% of people keep rolling dice coz number go up is just a computer game.

With each iteration, the casino crowd grows in size and ultimately more people get “onboarded” into crypto, there is always some new narrative that brings in additional people…the allure of “this is game changing”.

The reality is maybe it’s not *actually* game changing, and I think people get wiser to that with each iteration. However, whether it is game changing or not shouldn’t be conflated with price action, and imo the casino crowd is rather clinical at creating big speculative bubbles that results in crazy price action, because at the end of the day crypto is still tiny in the context of traditional markets.

Anyways, provided the overall market conditions hold up (and that is an important requirement), it feels like to me that AI is the next iteration of the hyper gambling movement. It has the same faces from the memecoin cycle, it is bringing in new people under the guise of “this is game changing tech”, AI bros have now realised they can actually monetise their work with crypto (just like how artists did in 2021), and it’s obviously a huge narrative in traditional stocks and also our everyday lives.

Is the tech actually game changing? I have no idea, I think there’s some cool shit and I think there’s some pointless shit; I doubt we are seeing the cutting edge in crypto…but I’m not gonna be the one to judge, and maybe it’s not even that relevant.

The point I’m trying to make is that it doesn’t have to be “good tech” for prices to go parabolic…think about the idea of the metaverse in 2021…sounded great on paper, everything mooned, and now no one gives a shit…but the important point is that everything mooned and you could have made a killing.

So after initially being skeptical I think AI has all the DNA to become the next bubble in crypto, who knows whether it lasts but I think we haven’t yet seen the true parabolic price action en masse…and it’s coming.

[olimpio @OlimpioCrypto]

If you ever make swaps via any DEX, this is for you.

Some tips to maximize return in DEX swaps:

1️⃣ Use a DEX Aggregator. 1inch is very good, but using swap.defillama (aggregator of aggregators) is better

2️⃣ After identifying the best one via DefiLlama, compare directly with their front ends. Sometimes final values differ.

3️⃣ MEV: protect against it. There are different alternatives to do this:

• Using Cowswap.
• Using 1inch Fusion instead of a regular trade.
• Using Flashbots Protect RPC (https://t.co/5Coglue1WV)

They all work in a similar way: your transaction is not broadcasted to the mempool directly, and thus it's protected against front-runs.

Sometimes you find the best rate in a DEX that does not offer MEV protection (for example Paraswap or Odos) and Flashbots RPC protects you in those cases.

It also allows you to save gas fees on failed transactions: if you set slippage too low (which you should in most cases) then if the transaction fails it does NOT get broadcasted if using Flashbots. Win-win-win.

4️⃣ Find where the best pools are, with the most liquidity. Get the token smart contract from CoinGecko and then paste it on @GeckoTerminal or @Dexscreener to see where the token is being LPed and trading in all of DeFi.

There are some rare cases where this might help discover a pool in a protocol that's not yet integrated with the aggregators.

Also, pay attention if the token was deployed in different chains - sometimes the token contracts are different, and they have fragmented liquidity amongst different chains.

5️⃣ Same DEX, different fees: a trade in Curve is usually expensive compared to a similar one in Uniswap. This is because of how Curve's smart contracts work.

> If you trade on Curve via 1inch, it's cheaper (gas-wise).

6️⃣ Try Gashawk (https://t.co/xWD7MuLJQE) to save fees on Ethereum transactions that are not time-sensitive. Like Flashbots, it's basically another RPC you add to your wallet.

After you confirm the transaction in your browser, it is not directly broadcasted to the chain, but placed on hold. They have a dashboard where you can monitor all pending transactions (and cancel/speed them up).

The idea of having this setup is that GH automatically broadcasts the transaction on their own once gwei is cheaper, allowing you to save money on fees.

How does it work? After you sign the transaction locally, the signed hash (valid and ready to be broadcasted) is stored on their servers. They only delay sending it to the chain until gwei is lower. Same as Flashbots, they don't access any keys or can tamper with the transaction.

7️⃣ Limit orders: modern DEX aggregators like Matcha, 1inch, and Cowswap let you place "limit orders" in DeFi.

How do they work? Similar to GasHawk - you sign the transaction on their front end, and they only broadcast it when/if the price you set is met. If not, they expire. You can also cancel them manually.

8️⃣ Which of these doesn't have a token? 🪂

DEX Aggregator airdrops are the easiest to qualify. Remember 1inch and Cowswap?

More trades+More volume=Bigger airdrop. There are a few ones that do not have a token yet:

• Matcha (though they are affiliated to 0x Protocol which already has a token, and they explicitly said "no token")
• Odos (supports multiple chains, like Polygon)
• Bebop (backed by Wintermute)
• DeFillama (although it's important to mention that there is no DefiLlama smart contract interaction here. If they wanted to take a snapshot, they couldn't do it since you interact directly with the other protocols. Something could change in the future)

⚡️