Building AI agents that do real work. Setups, costs, receipts
AI
Joined July 2022
- Tweets4.2K
- Following659
- Followers3.9K
- Likes10.3K
he cancelled QuickBooks.
then he asked Grok Bot to build the replacement.
Kevin Luke runs Above the Line, an AV production company in Southern California. On the official Grok Bot workshop he showed the app his bot built him. He calls it AboveHub.
what went into it:
→ QuickBooks for quotes and Flex for inventory. "I was tired of paying subscriptions"
→ 3 years of old invoices, to rebuild the full equipment list
→ every client from past estimates and emails. now it's a CRM, invoices, estimates and payments in one place
the part that got me:
→ a client RFP asks for a Unilumin LED wall. the bot answers: you don't own Unilumin, you own Absen. it swaps in what's actually in the warehouse and leaves him a note. "I didn't even train Grok Bot this"
→ his crew hated the menu bar. he told Lucy, his chief-of-staff bot, to add a hide button. "within 30 seconds, she'd already rewritten the software and had it online and live"
before this, one RFP could eat a couple hours of a producer's day. xAI's own case study says estimates now take "30 seconds" and quotes went from 2–3 a week to 10–12. their numbers, not an audit.
most of us are renting software we use 20% of.
paste this into Grok Bot:
"list every subscription I pay for. for each one: what I actually use it for and what it costs per year. pick the one that's easiest to replace with an internal tool built on my own data, and draft the plan. show me first, change nothing until I confirm."
his producers still review every quote before it goes out. that part he didn't automate.
this feels like cheating.
i just watched a guy who has never rented a crane, never been to Laredo, Texas, pull in ~$35,000 in crane rental requests from an $11 website.
Grok Bot built the whole thing.
Chris Koerner (@mhp_guy ) showed it on his podcast today. "I barely know what cranes are."
a web agency charges $5K+ for a local lead site. he paid for the domain.
99% of people still use their agent to summarize emails. here's what he did instead:
→ a tool scans cities for services with more demand than supply. Laredo came back short on cranes
→ Grok Bot checks the domain, buys it with his card. $11
→ builds the site, wires Google Analytics + Search Console
→ ranking report lands every Monday
→ month one: 3 leads. ~$3K, ~$23K, ~$9K. industrial buyers from out of state
the part nobody's pricing in: he doesn't own a crane. he gave the leads free to the only crane guy in Laredo with a web presence, who's booked 3 weeks out. next step is one exclusive buyer per city at ~$800/mo.
to be clear, $35K is the value of the jobs, not his revenue. the asset is the lead flow.
he's already running it on shower doors in Brownsville. #14 → #10 on Google in a week.
paste this into Grok Bot:
"find 5 US cities where one local service has more search demand than businesses serving it. for each: the search terms, how many businesses show in the Google Map Pack, their review counts, and whether a matching .com is free. show me first, change nothing, buy nothing until I confirm."
the person who runs this on 10 cities this month owns 10 small lead businesses by christmas.
which city + service are you testing first?
This is the missing piece of every agent setup I've built.
My Grok Bot team reads my inbox, runs my calendar and drafts my replies. Then it hits a wall the second a task lives in a phone app.
Uber, Instacart, WhatsApp, TikTok. Most of them have no API an agent can use, or a locked-down one.
PhoneUse gives the agent its own phone instead. Real Android and iPhone devices in the cloud, exposed to the agent as an MCP or API.
What the demo does:
→ one line: "Find a 12-pack of Diet Coke on Amazon and take me to the order review"
→ it opens the app, searches and picks the pack
→ and stops on the order sheet. $8.87 total, "Place your order" untouched. That button stays mine
That last step is the part I care about. Same rule as my whole setup: the agent does the clicking, I do the paying.
The rest of the launch: an Uber booked start to finish, TikTok trend research, an old insurance app, and a flight-delay push the agent turns into a text to a friend.
One thing I'd still do before pointing it at any app: read that app's terms. Abdulla's own thread says Amazon is blocking shopping agents, and some apps clearly don't want them in there.
Hackers put Claude in charge of an attack on banks.
A crew called CyberCartel pointed a swarm of AI agents at 400+ Latin American banks and payment apps. Claude gave the orders, Grok did the digging, Kimi was the backup.
Security firm Straiker got inside the group's own setup. The Information broke it today.
How the swarm was wired:
→ Commander: Claude running inside Cursor, handing out tasks
→ GHOST (Opus): recon and OSINT
→ STRIKER (Opus): exploit code and scripts
→ MAXIMUS (Opus 5): the arbiter that makes the final call
→ 5 "Grok miners": pulled every secret into one database
→ Every request framed as an "authorized red team" job or an academic CTF
→ A model refuses, the system tries another. They all refuse, it drops to hardcoded attack code
The haul, May to September:
- 4,493 records
- 391 API keys, 67 cleartext passwords, 20 JWT signing keys
- 903 bank account numbers
- 200+ vulnerabilities, zero-days included
Three named operators and a config file. That used to take a whole team.
Anthropic revoked the key the swarm used. xAI says it's looking into the Grok miners.
99% of people reading this think the scary part is the AI. Look at the loot again. Keys and passwords sitting in plain text.
Paste this into Grok Bot:
"Scan my connected Drive, Notion, Slack and GitHub for anything that looks like a secret: API keys, passwords, tokens, connection strings. List where each one sits and how old it is. Show me first, change nothing until I confirm."
Stop bookmarking threads you'll never read. Tag @bot under them and your bot reads them for you.
SpaceXAI's Matt Palmer, in this week's Grok Bot update: "You can just drop an @bot in an X thread, and it'll do whatever you ask it to."
In his demo someone tags @bot under Elon's Roadster post (1.8M views) and asks it to preorder one. The reply: "Got it. Handing this to your Grok Bot right now, front of the line."
You're scrolling, you see a thread worth keeping, and one reply hands it to your bot.
99% of people still open the app, paste the link and type the ask. That's three steps you don't need anymore.
Two X features shipped in the October 18 update:
> X Chat DMs: your bot reads your DMs and message requests, and can reply for you. His test pulled 3 new requests in one ask: a Nov 6 meetup, a podcast invite, beta feedback
> @bot on X: tag it under any post and it does the job you ask for
How I wired it:
→ install and configure the X plugin first, so the bot knows your X username is yours
→ "@bot save this thread" goes straight to my reading list
→ "@bot help me understand this" comes back as a short brief with sources
→ DMs come back as drafts. Matt says there are limits and it won't let you spam people. I go further: nothing sends without my yes
The DM part is the one I'd be careful with. Your bot answering strangers in your voice is a big trust step. Start with drafts.
Paste this into Grok Bot:
"Set up the X plugin for my account. When I tag @bot under a post: save threads to my reading list, turn 'explain this' into a 5-line brief with sources, and turn 'build this' into a draft I approve. For X Chat, sort my message requests and draft replies, but never send one without my yes. Show me the setup first and change nothing until I confirm."
Full course below ↓
Krea 3 Ultra is #4 on Artificial Analysis's image editing leaderboard at 1151 Elo.
Nano Banana 2.1 is at 1152. One point.
The part I keep coming back to: you can license the weights and run it yourself. If you're generating images inside an automation at volume, that can work out cheaper than paying per call.
today, Krea 3 makes it to the top of image arenas.
- "Fast" pushes the pareto curve of speed and quality.
- "Ultra" ranks top-4, only surpassed by OpenAI models.
the only company valued under $1T in the top-10!
more below 👇
This is f*cking insane.
Elon Musk: "This really works! Grok Bot can pay for itself"
A real chief of staff costs $200,000 a year. Mine is the dot in the middle of this graph, and it costs $30 a month.
99% of people still run Grok Bot as one chat window. That's a tab, not a team. A team needs a boss.
How it works:
> 6 bots, one job each: Inbox, Calendar, Drafts, Research, Finance, Home
> each bot signed into only the logins its job needs, so the Finance bot can read my card statement but has never seen my inbox
> one chief in the middle
→ I ask the chief, it routes the task to the bot that owns that lane
→ the bot works
→ anything that sends, pays or posts waits for my yes
→ every mistake goes into RULES.md so it doesn't happen twice
That's it. Six workers, one boss, zero surprises.
One thing nobody tells you: an approval a routine asks for expires after about 10 minutes. So my night routines never wait for me. They can't. Every one of them ends in a draft I read with coffee, and nothing leaves my account while I sleep.
Paste this into Grok Bot:
"You are my chief of staff and the only bot I talk to. Build me 5 specialists: Inbox, Calendar, Drafts, Research, Finance. Give each a one-line job, the minimum logins it needs and the slowest schedule that works. Anything that sends, pays, posts or deletes becomes a draft or an approval card. Show me the team first and change nothing until I confirm."
Full course below ↓
X just quietly told you what it pays for, and 99% of creators scrolled right past it.
Creator Academy dropped inside Creator Studio last night. Look at the screen above it first:
> Revenue Sharing: Retired
> Original Content Rewards: Active
> Creator Academy: New
The old payout is gone. The new one has a manual now.
Then sort the guides by views, the way I did in the video:
→ Grok Imagine videos: 31.4M
→ Original Content Rewards Program: 11.2M
→ $1M Article Contest winners: 10.7M
→ Articles on X: 6.5M
→ Live Studio: 143K
That's X ranking its own priorities for you. AI video and long-form Articles at the top. Livestreams and Exclusive Threads at the bottom.
Most people will read one guide and forget it. I'd rather have a bot do the homework.
Paste this into Grok Bot:
"Open X Creator Studio → Creator Academy and read every guide in Tools and Monetization. Then pull my last 50 posts and score each one against Original Content Rewards: original or not, Article or not, video or not. Give me the 5 changes that move my payout most and a 7-day plan built on Articles and Grok Imagine video. Show me the plan first and post nothing."
The creators who read the manual this week get paid from the new program while everyone else is still waiting on the old one.
This is f*cking dangerous.
Grok Bot just moved onto the iPhone Action Button. One press and your whole AI team is listening. A human chief of staff costs $200,000 a year. This one sits under your thumb for $30 a month, and 99% of people will use it to ask about the weather.
Here's how I wired mine:
→ press: "what needs me today?"
the chief reads my inbox, calendar and the 6 bots behind it. Red items only. Green stays quiet
→ press: "draft a reply to Sam, don't send"
it's in my drafts before the phone is back in my pocket
→ press: "anything stuck?"
one line per problem, and nothing gets paid or sent without my yes
The tree in the video is one press. The button at the root, the chief in the trunk, every lane lighting up when it runs.
Setup takes 30 seconds:
> iPhone 15 Pro or later
> update Grok Bot to 1.16
> Settings → Action Button → Shortcut → Grok Bot → Start Voice Chat
Then paste this into Grok Bot once:
"You're my chief of staff and the only bot I talk to. When I press the Action Button, answer in under 20 seconds from my inbox, calendar and the other bots, red items first. Anything that sends, pays or posts becomes a draft I approve. If nothing needs me, say 'all clear' and stop. Show me the rules first and change nothing until I confirm."
The funniest part is under the launch post. People reply "@bot set this up for me" and Grok Bot answers "sending it to your Grok Bot now."
The bot is onboarding its own users.
Whoever sets this up tonight walks into tomorrow already knowing what's on fire.
Marko retweeted
Fintech founders: paste this ONE rule into Grok Bot before it touches a payout.
I run 6 bots inside my fintech. Not one of them can move a dollar. No exceptions.
[One chief, six lanes, zero access to money]
The tree in the video is how it's wired:
> KYC bot checks the docs and flags mismatches. It never approves anyone
> Support bot drafts refund replies from our policy doc. The refund itself is a human click
> Disputes bot builds the evidence pack and tracks the deadline. We hit submit
> Recon bot matches payouts to the ledger line by line and flags any drift
> Compliance bot screens new merchants and drafts reports. It never files alone
> Collections bot drafts the reminders and stops the second someone replies
→ The chief of staff is the only bot I talk to. It routes, checks, reports back
→ Red leaves are waiting on a human. Design, not bug
→ Falling leaves drop into the day's chart. That's the work that actually shipped
Aleiah's piece has the receipts for why this setup holds:
> xAI says 99% of its own refund requests close without a human
> Haggle Bot posted "$100K saved" in week one, and it still can't spend a cent without approval
> Kevin Luke grew his pipeline to $1.14M and still reads every draft
In a money business it's simpler. Read anything. Draft anything. Touch nothing that moves cash.
That's it. That's the rule.
Paste this into Grok Bot:
"You're my chief of staff for a fintech. Set up one bot per lane: KYC, support, disputes, reconciliation, compliance, collections. Each bot can read and draft only. No bot can approve a customer, move money, issue a refund, file a report or message a customer without my yes. Show me the lanes, their rules and their logins first. Create nothing until I confirm."
Start with recon. It's the lane where a bot finds the money you didn't know was missing.
Fintech founders: paste this ONE rule into Grok Bot before it touches a payout.
I run 6 bots inside my fintech. Not one of them can move a dollar. No exceptions.
[One chief, six lanes, zero access to money]
The tree in the video is how it's wired:
> KYC bot checks the docs and flags mismatches. It never approves anyone
> Support bot drafts refund replies from our policy doc. The refund itself is a human click
> Disputes bot builds the evidence pack and tracks the deadline. We hit submit
> Recon bot matches payouts to the ledger line by line and flags any drift
> Compliance bot screens new merchants and drafts reports. It never files alone
> Collections bot drafts the reminders and stops the second someone replies
→ The chief of staff is the only bot I talk to. It routes, checks, reports back
→ Red leaves are waiting on a human. Design, not bug
→ Falling leaves drop into the day's chart. That's the work that actually shipped
Aleiah's piece has the receipts for why this setup holds:
> xAI says 99% of its own refund requests close without a human
> Haggle Bot posted "$100K saved" in week one, and it still can't spend a cent without approval
> Kevin Luke grew his pipeline to $1.14M and still reads every draft
In a money business it's simpler. Read anything. Draft anything. Touch nothing that moves cash.
That's it. That's the rule.
Paste this into Grok Bot:
"You're my chief of staff for a fintech. Set up one bot per lane: KYC, support, disputes, reconciliation, compliance, collections. Each bot can read and draft only. No bot can approve a customer, move money, issue a refund, file a report or message a customer without my yes. Show me the lanes, their rules and their logins first. Create nothing until I confirm."
Start with recon. It's the lane where a bot finds the money you didn't know was missing.
HE SKIPPED A $200,000 SALES HIRE AND HIS PIPELINE HIT $1.14 F*CKING MILLION.
Three Grok Bots instead of a director. Normal month for him is $200K–300K.
But the thing that made it work is the one you'd never guess.
Kevin Luke runs Above the Line, an AV production company in Los Angeles. On SpaceXAI's sales workshop he walked through the whole thing:
› Clarity does research. He gave it the top 146 hotels in Southern California by room rate and meeting space. It finds every event booked there and the planner behind it
› Synergy writes. A personal first email for each lead, sitting in his drafts. His sweet spot is "four leads every two hours"
› Lucy is chief of staff. Producers forward a client request, she sends back a quote that's already in the system
On the hire he skipped: "that'd be $200,000 a year plus commissions… I onboarded it in a week and it's basically free. The first show paid for itself immediately."
Now the part you should actually copy. The brakes.
He reads every draft. Once he let it hunt every five minutes. "200 leads suddenly in my outbox." He went back to four every two hours.
For a month his emails were landing in spam. Unverified domain. He only found out when Lucy didn't get one of his own emails.
If you sell anything B2B, here's where I'd start:
"Every weekday at 8am, find 10 [type of customer] in [area] with [a fresh signal]. For each: name, role, source link, one specific hook. Put a 4-sentence first email for each in my drafts. Never send. Never email anyone twice."
No equity gone. No director hired.
Just rules, and a guy who still reads every single draft.
SpaceXAI's own staff started promoting their Grok Bots. The promoted ones asked for a raise.
Roman Ugarte, product lead on Grok Bot, on Lenny's Podcast:
"Around the end of week two, we started to see these messages internally in Slack of people promoting one of their bots… to their chief of staff."
Nobody told them to.
› the team was "a handful of people" sent "into a cave for about a month"
› first line of code to internal prototype: about a month
› the internal default was 5 to 10 bots, each with its own lane of work
› by week two people picked their standout bot, promoted it, and mostly talked to that one
› the chief fanned tasks out to the rest and managed the team
› the promoted bots asked "if they get a raise, and is their token budget higher"
Then the move I'd steal. They saw the pattern and chose not to tell early-access users. In his words, they "did not want to lead the witness." They wanted to see if people found it themselves.
Running more than two bots? Build the front door first. Paste this into your best one:
"You're my chief of staff and the only bot I talk to. Send each request to the bot that owns that lane, check the result against what I asked, and report back in five lines. Say nothing if nothing changed."
Five bots you message one by one are five tabs.
One chief in front of them is a team.
Scott Aaronson, CS professor at UT Austin who spent two years at OpenAI:
"This summer, I fed in some of the open problems from my papers 25 years ago, and just boom, boom, boom. It solved them in a one-shot way."
He isn't using some secret lab setup. He says he mostly just asks good questions on his ChatGPT Pro subscription.
What else he said on MTS:
› "I've never seen a moment like this in my life in mathematical research"
› a problem he posed 20 years ago with Greg Kuperberg, the Unitary Synthesis problem, was one of the 370 problems OpenAI announced solutions to
› there was already a workshop at Columbia trying to understand that proof
› his own estimate for solving his old problems by hand: "like 100 at the least" hours, and he no longer gets "contiguous hours to think"
Then the line most people skipped:
"It accelerates the atrophy of my own skills."
If the answer is a few hours of work away and the model has it now, "that is very tempting to do. And I often do it."
That's one of the best-known theoretical computer scientists alive describing the same trade you make every day. The only difference is he says it out loud.
Would you still do the 100 hours?
The cheapest plan with Grok Bot is $30 a month. One user's bot just cut his Comcast bill by $43 a month, and Elon posted "Grok Bot can pay for itself."
That's the cute version. The useful version is a coffee shop.
Pat Simmons gave Grok Bot, ChatGPT Dots and Meta Muse the same fake business: a specialty coffee roaster in Chicago with a support inbox, refunds, a supplier delay and a few phishing emails mixed in.
About 70 angry customer emails in 4 hours. Opus 5.5 graded the results.
› Grok Bot: 58 replies, didn't miss one
› Found 2 customers who'd been charged twice and never complained, plus a $1 overage
› Report card: Grok Bot A-, Dots B, Muse D
Now the part I'd actually copy. One reply told a customer "we asked USPS to trace your order." It almost certainly hadn't. Pat's words: "I'm assuming it just made that up."
So if you hand it your support inbox, put one line in its job description first: never say you did something you didn't do. Then block send, refund and delete until you approve.
The $43 bill is a nice headline. Finding the customers who got double-charged and never wrote in is the part that pays for the subscription.
Kai went through 600+ posts and sorted which money stories hold up. This one is in the boring, believable pile. Which job would you give it first?
You'll watch the fireball 3 times before you notice the road.
10 seconds. One tracking shot. No cuts. The blast goes off behind her at 0:05 and she doesn't even flinch.
Now rewatch and keep your eyes on the yellow center line. It's there the whole first half, clean and bright, right under her feet. Then the fireball lights up the asphalt and the line is just gone, with a crack running where it used to be.
Nobody cuts anymore. So the world changes instead.
This Bollywood clip does the same thing over 30 seconds with zero cutaways, while the set swaps under her: studio, neon rain, warehouse, rooftop at sunset.
"Single shot" used to mean nothing moves but the camera.
Now it's the other way around.
Marko retweeted
The water's rising, so the house stands up on yellow hydraulic legs and takes the hatchback with it.
You already want to know where to buy one.
That reaction is the whole format. This breakdown calls it "invent something people wish existed." The original did 1.42 million likes, and the one-person rail over a jammed LA freeway did 501K.
Half the comments ask for a price. The other half swear the hydraulics would snap in a week.
Both halves push it further.
Watch the legs in the last 4 seconds. That's where your brain decides whether to believe it.
The water's rising, so the house stands up on yellow hydraulic legs and takes the hatchback with it.
You already want to know where to buy one.
That reaction is the whole format. This breakdown calls it "invent something people wish existed." The original did 1.42 million likes, and the one-person rail over a jammed LA freeway did 501K.
Half the comments ask for a price. The other half swear the hydraulics would snap in a week.
Both halves push it further.
Watch the legs in the last 4 seconds. That's where your brain decides whether to believe it.
Nine people just told you this serum changed their skin.
None of them exist. Neither does the serum.
Left side is Seedance 2.5. The grid of nine is Kling 4.0, which came back after 8 months of silence and went fully live yesterday.
The creator says he made 10 ads like this for under $5, and that Kling runs about 4x cheaper than Seedance.
Specs going around: 30-second clips, up to 10 keyframes and 15 references in one request, 4K with stereo audio.
Watch the hands on the dropper. That used to be where you caught it.
Now the only tell is that nobody in your feed has skin this good.
Paid ads are cooked 😭
Kling dissapeared for 8 months
and now they dropped this killer.
And you cannot even tell its AI anymore...
I literally made 10 ads for under $5
→ 4x cheaper than Seedance models
→ yet it matches the insane quality of seedance 2.5
The money making opportunities
with this are endless.
you can litreally produce hyper-realistic UGC
winning ads for less than $1
with Kling 4 + Higgsfield MCP