Vegas, the knowledge bases, and the four things Tim asked for
Friday, August 7, 2026 · 8:16–9:02am · Paul & Tim · phone call, 46 minutes
Written up from the call. Melissa wasn't on it — Tim, if you can brief her, the four requests
in part two are the part that matters. Everything below is what was actually said, not a plan.
“Best talk I've ever had with Tim. The whole time I've known him — this morning was the best talk.
And we just talked on the phone. We didn't even do a Zoom.”Paul, right after the call
Part one · Tim's report
Ai4 Las Vegas — the scale changed everything
12,000attendees
~500vendors
3 daystalks & floor
1
It stopped being a niche show
Tim has been to several. It used to run 50 to 150 vendors, and you couldn't find a talk on
construction or anything adjacent to our work. This year there were talks on everything.
The problem that creates
“There's so many software companies trying to sell AI infrastructure software. I don't know how a
CIO would ever figure out what they need.” That confusion is the opening — it's the same reason a
home builder can't evaluate five marketing vendors.
2
A whole software layer just to watch the agents
Beyond tools for building custom agents, there's now an entire category monitoring them:
security, hallucination, who has access, how they collaborate.
Worth noting because it's the thing we've been doing by hand all week — checking one
agent's work with another. The market has decided that's a product category.
3
Every talk is being filmed, and Tim has access
The full conference goes onto the event app within a few weeks. Attendees get it;
nobody else does.
Why that matters
That's proprietary material no agent could otherwise reach. Paul: getting it into the knowledge base is
trivial — the only requirement is access.
Part two · the actual asks
Four things Tim wants the system to do
These came out of the conversation unprompted, and they're all buildable. This is the
agenda for Monday.
1. Watch the people who matter on LinkedIn, every day
Tim follows people in the magnetics world who post studies and thought-provoking material. He doesn't
read it daily and doesn't want to. He wants an agent that does — and that surfaces the adjacent
people he's never heard of.
Known obstacle: LinkedIn blocks bots. Paul's position on the call: there's a way around it,
it's doable.
“I'd really just like an agent to figure out eventually everybody in the world that's posting magnetics.”
2. Turn what it finds into our own articles
Not just monitoring. Which of these should we repost — and which should spur our own piece about
PolyMagnet's role in whatever they're discussing?
Tim's phrase: “a self-fulfilling engine. Every day it comes up with new ideas.”
3. Feed it the conference
All three days of Ai4 talks, once the videos land on the app. Read and listened to, into the
knowledge base.
Needs from Tim: access to the app or the video library.
4. An agent whose job is to interrogate you
The strongest idea on the call. The system can't know what you read, attended or were sent. So a
dedicated agent should probe at each review: did you go to a conference? Is there proprietary material
— written, video, a site — that should go into the database?
“You wouldn't think of it. You might go — oh, I read that really cool magnetics article by
so-and-so. It doesn't know you did that. It wouldn't know unless it had a dialogue with you at that
management meeting.”
Part three · where the product actually is
The knowledge bases are built
Tim's challenge on the call was fair: most agentic AI does rote repetition — same file,
same time, same cells — and very little of it genuinely generates new ideas. Here's the answer to that.
85engines live
76knowledge bases
3,946verified companies
4,284buyer questions
816sourced citations
88magnetics companies
4
The taxonomy came from the Austin Business Journal
Their Book of Lists is a worked-out map of a market — home building, commercial real
estate, residential, the whole thing. Every category became a knowledge base.
Why it travels
ABJ runs the same model in hundreds of cities. The taxonomy is a starting point we can replicate market
by market, not an Austin-only asset.
5
Everything was verified, and that found real errors
Each company has a working URL, a role and a geography. Every industry claim now carries a
source and a date. 154 were deliberately left blank rather than dressed up with a company homepage
pretending to be a citation.
What the checking caught
Insurance described as hardening — it has fallen for eight straight quarters. Two knowledge
bases had it backwards.
An IRS ruling cited that was never entered; the case was dismissed.
Third-party cookies described as dead. Google reversed course.
An ABB robotics spin-off that never happened — the division went to SoftBank.
Dead links that would have shipped live: a folded Austin soccer club whose domain now serves a
gambling site, twelve closed coworking spaces.
This is the difference between a big knowledge base and one you can point at a client.
None of it surfaced from counting how much research we had.
Go and look · in priority order
Everything that changed while you were in Vegas
Open the first one. If you only click three, make it the first three.
1 · The engine itself
automarketingengine.com — the mothership.
Run a site through it. This is the one that matters; everything below is a specialised version of it.
2
Construction & building — your world
Ten separate engines, each with its own knowledge base. Home building is the deepest on
the whole network.
What that actually means
A builder's site run through this doesn't get generic marketing copy. It gets copy that knows what a MUD
tax does to a payment, why an option period matters in Texas, what an impervious-cover limit is, and
which of the 62 builders it is actually competing against in that county.
3
Magnetics
magnetics.wholereach.com —
88 companies, mine to finished magnet: rare-earth mining and separation, alloy and powder, global
NdFeB and SmCo makers, magnetising equipment, distributors, couplings and Halbach arrays.
It now also carries 52 specification concepts — remanence, coercivity, BHmax,
permeance coefficient, the N/M/H/SH/UH/EH grade suffixes, UN3506 air-shipping rules — and
60 buyer questions, so it speaks like an engineer rather than a brochure. Related:
manufacturers.
One live finding from the sourcing pass
The DoD ban on Chinese-origin NdFeB and SmCo takes effect 1 January 2027. USA Rare Earth is now
USARE with magnet production in Stillwater, and MP Materials has a magnetics division in Independence,
Texas. That is a dated, sourced supply-chain shift sitting in the knowledge base right now.
4 · The whole map
lists.wholereach.com — every industry segment
we run, sortable. Click any row to see what that engine actually knows. 85 engines, 76 knowledge bases.
The taxonomy came from the Austin Business Journal's category lists — their worked-out map of a
market. ABJ runs the same model in hundreds of cities, so this is replicable market by market rather than
Austin-only.
The part you actually asked about
How it keeps up — and where your LinkedIn idea fits
Tim, your challenge was that most agentic AI does rote repetition and very little of it
genuinely comes up with new things. Here is exactly what runs on its own today, and the one gap.
●
Every Monday, 7am — the whole network re-gathers
All 76 knowledge bases re-fetch every company in them. A firm that has been acquired,
renamed or shut down shows up as changed that morning, not months later.
●
The 1st of each month — a decay audit
Checks all 3,946 company URLs and separates moved, parked and dead,
then pushes the list to a phone.
First run found 93 real problems
including a folded Austin soccer club whose domain now serves a gambling site, and twelve closed
coworking spaces. All would have shipped as live links in client content.
●
Every industry claim carries a date
816 of them now have a source and an as_of stamp. That is what makes staleness
detectable at all — and it is how we caught insurance being described as hardening when it has
fallen for eight straight quarters, and an ABB spin-off that never happened.
The gap is precisely what you asked for
Weekly re-gather and a monthly audit is a heartbeat — it keeps the knowledge base from
rotting. What it is not is daily attention on named people.
Your request — follow the magnetics voices on LinkedIn every day, surface the adjacent ones
neither of us has heard of — is the layer that turns the heartbeat into the self-fulfilling engine
you described.
The knowledge base is the part that makes it worth doing. An agent that reads a post and knows the
88 companies, the 52 spec terms and the DoD deadline can judge whether that post matters —
and draft the PolyMagnet piece it should spur. Without that, it is just a clipping service.
LinkedIn blocking bots is the only real obstacle, and it is a solvable one.
Same shape for construction: name the ten people whose posts you would
read if you had time, and that becomes the daily layer on top of a 62-builder knowledge base.
Part four · how we build from here
Multi-agent, multi-person, multi-model
6
Buzz — shared sessions, three weeks old
From Jack Dorsey's team. It turns a Claude session into something several people can join:
Paul building the interface, Tim watching the questions and answers in real time, other models in the
same room. Ships with agents that oversee and check each other, and you can spawn more.
Paul is running several instances and taking it in steps — not going all out until
the security side is settled. Team has been working in Claude together for three to four years; this is
the first tool that makes the group version work.
7
What forty agents at once actually taught us
The knowledge bases were built by roughly forty research agents running in parallel, each
owning one industry.
The surprise
They were consistently better than their instructions — correcting a firm we placed in Austin that
was in Massachusetts, another in Scranton, a trade body that doesn't exist, an MBA programme we thought
had closed. Roughly a quarter of what came back was a correction rather than an answer.
Two things that bit
Agents editing the same file overwrote each other — fixed by giving each its own files. And four
times we added good research that nothing in the software actually read. New rule: wire it up and prove
it reaches the page before researching more.
Agreed
Next steps out of the call
Tim — write up your genuine Ai4 takeaways and send them. Paul wants to fold them into the
process. Also: access to the conference videos when they post.
Monday — call to go through the individual agents one by one and what each actually does.
Paul drives back from Hot Springs Sunday, so Monday is the basics; Tuesday is full engagement.
Open idea — a speaking slot at next year's Ai4 to roll out the marketing department. Paul's
view: shouldn't take a year.
Melissa: anything from Vegas that contradicts the above, say so before Monday.
Worth knowing, unresolved
Austin's numbers have turned. The sourcing pass found metro wage growth flat and
information-sector employment down two years running. Doesn't break the strategy — a soft market
still needs marketing — but "grow with the boom" is a weaker line than it was.
LinkedIn monitoring is the gating technical problem for requests 1 and 2. Everything else on
the list is straightforward.
Nobody has read a full report as a customer would. The machinery works; the writing hasn't
been judged. That's a reading job, and it's Tim's to do.