I write about AI a year before the takes catch up. Case studies, not commentary.
- Tweets 595
- Following 137
- Followers 369
- Likes 807
SpaceXAI runs Grok Bot on itself.
One internal setup: a chief-of-staff bot directing specialists - recruiting, expenses, bug fixes. The bots share a thread, pass work between each other and coordinate in a group chat.
Each one gets its own cloud machine, so the work continues after you close the laptop.
Most people run a single bot and stop there. The product was built for the other case.
Below: how to structure a bot team for your own repo.
2
8
1,412
Ex-Google engineer reveals how to build AI agent loops, harnesses, LLM ops, and evals in 19 minutes.
Trace → evaluate → diagnose → fix → ship → repeat.
That loop is how agents self-improve over time.
Agentic loops + harness + memory + evals - that's the senior engineer stack.
This is better than $500 paid courses on the same topic, explained in under 20 minutes.
Watch it, then save the framework below.
2
1
6
327
Diogo Almeida co-created RLHF at OpenAI. Two years later he built a model that doesn't use it.
His company, TypeSafe AI, shipped Jev. It doesn't write text at all - it picks one of the options you defined and returns a calibrated decision.
That's why it can't hallucinate. There is nothing to invent.
Company numbers: 70-500ms responses, $42 per billion input tokens, output free.
Independent test: Vercel swapped its safety classifier over and got results 5-18x faster.
In a 36-min talk he explains why chat-tuned models were the wrong shape for automation.
Below: what building on this actually looks like.
1
1
1
236
Ex-Google engineer reveals how to build AI agent loops, harnesses, LLM ops, and evals in 19 minutes.
Trace → evaluate → diagnose → fix → ship → repeat.
That loop is how agents self-improve over time.
Agentic loops + harness + memory + evals - that's the senior engineer stack.
This is better than $500 paid courses on the same topic, explained in under 20 minutes.
Watch it, then save the framework below.
1
7
23
4,639
Anthropic released 5 workshops on building self-improving agentic systems from scratch:
00:00 - Ship your first Claude agent
36:44 - Self-improving agents (tools, skills)
1:21:25 - Build memory for Claude agents
1:49:47 - Set up a proactive agent
2:11:31 - Make your agent autonomous
These 3 hours of Claude workshops will replace 20 paid agentic courses.
Watch today, then read the article below on how to build a self-improving agentic system with Fable 5.1.
4
10
54
11,357
Andrew Ng just released a 1-hour course on building agentic knowledge Graphs from scratch:
00:00 - Introduction to agentic knowledge Graphs
03:07 - Construction of agentic Graphs
14:00 - Architecture of multi-agent systems
23:00 - Building agentic graphs with Google ADK
01:06:03 - Why Graphs are the future of agentic AI
Worth more than 10 articles on loop engineering.
Watch it today, then read how to become a graph engineer in the article below.
3
71
328
44,350
Google just dropped a 1-hour course on agentic engineering from scratch:
00:00 - How to build your first AI agent
08:24 - Build agent memory (short, persistent, long)
28:34 - Agentic loops, long-running AI agents
40:04 - How to build MCP (MCP vs API)
1:00:22 - Multi-agentic systems
This 1-hour watch will replace 10 paid agentic courses on the internet.
Watch it today, then read how to build a self-improving agentic system in the article below.
4
20
86
16,768
SpaceXAI engineer Lingxi Li builds Grok Bot using Grok Bot.
He went from 15 cloud agents he managed by hand to 200+ running through five named bots. A separate ops bot writes the postmortems. Every 30 minutes they sync progress into one shared Notion database.
The work continues while he's asleep or in meetings.
In a 30-min workshop he walks through the whole setup.
Below: how to build the same agent team for your own repo ↓
2
15
1,848
Most people use GrokBot like a search engine.
A SpaceXAI engineer (ex-Cursor) runs 20+ agents in a loop: Chief of Staff, GTM, Engineering, Marketing — plus 15 workers under them.
He broke down the whole architecture in a 1-hour workshop. Free.
Watch it before you pay $500 for an "agentic engineering" course.
Article on getting 100% out of GrokBot below 👇
4
1
17
1,796
The AI in this scam filter is told 50% of liquidity is burned or locked on every Solana token.
Nobody measured that. It's a placeholder in the code:
lpBurnPercent = 0.5; // TODO
AEGIS sends Grok the six checks and a safety score, then asks for a verdict: BUY, AVOID or CAUTION.
The score has a hidden input. The memory layer adds +10 for any deployer with 5 scanned launches who didn't renounce, because nothing in the repo ever marks a token rugged.
I ran that code. Output:
"trusted deployer: 0/5 tokens still alive"
Grok gets the +10 inside the score. It never gets the reason.
In this setup the model can't check the numbers. It can only explain them, so it analyzes a TODO and a bug as if they were data.
Caveat: Grok is optional, the project is v0.2.0, and I read the prompt but didn't run the Grok call. The fix isn't a smarter model. It's not sending numbers nobody measured.
Your scanner has never once remembered a scammer.
You checked the mint, the LP, the holders, and the deployer that rugged you last month passed every check the month before too.
I built one that remembers, straight from solana, base, and robinhood chain.
AEGIS, the scam filter with a second brain / github.com/andreysuperiorgit…
Paste any address, three seconds, a score from 0 to 100, every deployer AEGIS has seen kept on your disk.
I just shipped the second brain into it, the six checks tell you what a token is right now, the brain tells you what the people behind it have done before.
What it does before you click buy:
> Runs six on-chain checks, mint, freeze, top 10 concentration, bundle detection, LP lock, metadata
> Writes every scan to a local SQLite file, deployer, wallets, verdict, status live or rugged
> Adjusts the score by plus or minus 25, a deployer with 3 rugs starts at minus 15 before any check
> Refuses to trust a clean contract from a dirty wallet, known-ruggers in first buys knock 10 off
> Why this is the smart part:
A competent rugger passes six checks every time, that is the point of being competent, the contract is a snapshot and the snapshot always looks clean.
What they cannot hide is the pattern across launches, same funding wallet, same three bot wallets in the first block, and the second brain reads that pattern before the next victim clicks buy.
Runs on your machine, no cloud, no account, gets sharper the longer you run it.
1
2
1,256
Senior Forward Deployed Engineers at Anthropic can clear $785K/year in total comp
In a 20-min talk, Anthropic's Kevin Bai (ex-Palantir, first FDE at Rippling) explains why the role exists:
- you're not selling software or time. you're selling an outcome. "a leader in oil and gas cares about throughput, not how the data is organized"
- FDE fits one case only: technical product + non-technical buyer. everything else runs on DevRel or sales
- every AI platform is now agentic and customizable, so every AI company is in that case. "your customers have no idea what you actually do"
- the payoff, by his numbers: Palantir ~$4M average contract value, ServiceNow $1.2M, Workday ~$600K. no other public SaaS cracks $500K
watch, then read the guide below ↓
3
5
1,447
You're not supposed to prompt Claude.
You're supposed to build a system that prompts itself.
That's how one Anthropic engineer puts it.
In 55 minutes she shows how Anthropic builds agents that remember, fix their own mistakes and improve with every run - using graphs.
More specific than any paid agents course I've seen.
Watch it, then read the guide below on building the same system with graphs.
2
7
1,267
Most people use Claude Code like Google.
One Anthropic engineer runs 100+ agents in a loop:
"I have a Chief agent, PM agents - they manage the whole team"
In a 30-min workshop he showed how to get the most out of Claude Code without burning tokens
Skip the next $500 vibe-coding course and watch this instead
Then read how to build self-improving agent systems with Fable in the article below
4
7
1,544
xAI engineer (ex-Cursor):
"right now I'm running 10-20 GrokBot agents that automate 90% of my routine
i have a Chief of Staff agent. He knows about all my other bots and manages everything"
in a 50-minute podcast he walked through how to build a team of agents that runs 24/7
more useful than most $500 agentic engineering courses
watch it, then read the full breakdown on building a Grok agent team from scratch below
3
10
1,436
They asked the CEO of Anthropic if he would hand his company to the government.
He said yes.
Not as a joke. Not as a hypothetical. He said the right combination of governments should have it.
Then it got worse.
He said people constantly ask him why a private company is building this instead of a state, and that the strangest thing about the question is that he agrees with them.
His word: uncomfortable.
This is the man running the company. Saying out loud that he is not sure he should be running it.
Hours earlier he published a three step plan to slow the whole industry down.
Step one is the only one anyone signed: outside evaluators get permanent, employee level access inside Anthropic. Badges, laptops, the same permissions as internal risk staff, and the right to publish without his approval. Altman said OpenAI would match it.
Steps two and three are agreements between labs, then with China.
Neither exists.
He rejects banning the technology, because China builds it regardless. So the slowdown only fits inside the margin of the US lead, and no one can measure that margin.
Asked whether a real speed limit with China is possible, he said he does not know.
Ok this is starting to feel like a f*cking disaster.
The CEO of Anthropic just published an article admitting AI is already building the next generation of AI by itself.
He says within 6 to 12 months a rogue swarm could take over the entire internet and cause hundreds of billions of dollars in damage.
And what makes it scarier, Elon Musk just backed up everything Dario said.
All of this dropping just days after Jacob Coxon went viral with his warning about AI and the extinction of humanity.
Tell me this timing isn't strange.
20
10
49
29,559
Four accounts posted the same Lauren Tan quote this week. The quote contains three different numbers.
One version: 15+ agents, 3 managers, 11 workers.
Another: 20+ agents, 3 managers, 16 workers.
A third: 10 to 20 agents.
Nobody moved the quotation marks. They just moved the count.
What the workshop actually shows is 10 to 20 bots under one Chief of Staff routing work, and a claim of roughly 90% of her routine automated. That last figure came from company demos during launch week, so it is a best case, not a measurement.
The reproducible part was never the headcount. It is the architecture - a coordinator that holds context, specialists that do not, and a human gate before merge.
The number got copied. The structure got ignored.
1
5
1,653
SHE RUNS 20+ AI AGENTS. YOU RUN ONE.
99% of people using GrokBot are touching 1% of its power.
They run a single agent. No loop. No graph.
Lauren Tan, engineer at SpaceXAI:
"I run a team of 20+ GrokBot agents, fully autonomous. A Chief of Staff agent, a PM agent, and 20+ workers. That's the new engineering stack."
The shift is not the model. It's the org chart.
One agent answers a prompt. A graph of agents holds a goal, splits it, and checks its own work.
In one hour she builds that team from scratch.
This is worth more than a $500 agentic engineering course.
Watch the workshop today, then read how to build your own fleet in the article below.
3
1
15
3,210
Codex now ships the things people were building custom frameworks for.
One hour covers all of it:
00:36 agents and subagents, and what each spawn costs you in context
11:34 prompting patterns that survive more than one turn
20:06 /goal for runs that continue after the first answer
30:03 skills: the reusable playbook layer
41:28 /automations, scheduled runs that spawn their own subagents
The 30:03 mark is the part that pays back the hour. Skills sit in a
global folder other agents read too, so what you write for Codex
works outside Codex.
Caveat: /goal without a verifiable completion condition just burns
tokens quietly for hours. The loop needs a stop test, not a wish.
Full breakdown in the article below.
1
14
1,783
An Anthropic engineer put out a 1-hour breakdown of how agentic systems actually run in production:
00:19 - what Claude Code does under the hood
14:20 - long-horizon agents, CLAUDE.md and Plan mode
23:38 - skill design patterns for agents
44:01 - building agents and subagents
59:34 - self-improving loops and graphs
The last section is the one most people skip and the one that matters: an agent that logs its own failures and rewrites its own instructions stops needing you in the loop.
Full write-up on loops and graphs below.
1
1
9
1,693
a SpaceXAI engineer runs 20 agents at once. he talks to 2 of them.
the rest report to a Chief of Staff agent and a PM agent. he only reviews output.
"99% of people use GrokBot like Google. 1% are building teams of self-learning agents"
20-minute workshop below shows the exact setup: cloud agents, loops, delegation between them.
this is the $500 agentic engineering course, for free.
article on running GrokBot at full power is in the next post.
1
13
1,029