BlvkWare.
What we actually build
Everyone can show you a chat box. This is the whole machine.
A slide deck, a call script, a spreadsheet and a strategy memo went in. What came
out runs an eight-stage business process, refuses to do the three things that would get its owner
sued, and tests itself 203 times. Two pieces of it are running on this page: the part that
reads, and the part that shows you what happened. Both work. Use them.
4documents in
8stages enforced
203self-checks, all passing
6things it refuses to do
2live demos below
An agent is three things, and two of them are invisible
Most of what gets sold as an AI agent is the first column below and nothing else. That is why
it demos beautifully and then does something expensive in week three.
01
Something that decides
Reads what actually arrived, works out the next step, and knows when a case is outside
what it should decide alone.
The agent
02
Something that says no
The rules it operates inside. What it may touch, what it must evidence first, and the
things it is never allowed to do regardless of how confident it is.
This page, demo one
03
Something that shows you
What it did, what it is waiting on, and the part nobody ships: what it refused to do,
and why.
This page, demo two
Being precise about the word, since we sell it. Our
definition of an agent is that it handles a case nobody
anticipated when it was built. The system on this page does not do that, and we are not going
to call it an agent to make a page read better. It contains no language model at all. It is
column two: the machinery an agent operates inside, built so that the unanticipated case gets
refused rather than guessed at.
We build both. Column two is the one almost nobody shows you, because it is the one that is
hard, and because it is difficult to make caution look impressive in a screenshot. So here it
is running instead.
Demo one: the part that reads
A seller sends whatever they have: a maintenance-system export, a forwarded email, six
lines typed on a phone in a warehouse. This turns that into something a buyer can actually price,
and flags every line it could not identify rather than filling in a plausible guess.
This is the real engine: the same 94 patterns across 25 manufacturers the live system uses,
embedded from its own data files. Paste anything, or use the buttons.
Nothing leaves the page. The line it cannot identify is the important one:
it comes back marked unverified with a request for a nameplate photograph. A wrong part
number sent to a buyer is worse than no part number at all, so the system prefers the
embarrassment of asking over the cost of guessing.
The six things it refuses to do
Anyone can generate a codebase now. What is scarce is knowing what the software must
refuse. Each of these is enforced in code and has a test that fails if somebody later
takes it out. This is what we mean when we say our agents are built differently. Not cleverer.
Fenced.
01It will not sell to a buyer who never agreed to pay
All 20 buyers ship as published interest, meaning we watched them advertise,
rather than as customers. Nothing routes to them until a human records a confirmed rule, a fee
basis and an expiry date. On a fresh install the system can source nothing at all.
The plan said “no sourcing until buyer demand exists.” A
document says that once. Software says it every time, to everybody, including the founder
at 2am.
02One bid per buyer, and never the word “auction”
Texas exempts a sale by sealed bid from auctioneer licensing only when no bid can be
revised after opening. Mississippi and Louisiana carry comparable exemptions. So a second
number from the same buyer is refused unless a correction reason is stated, and the
correction is disclosed to the seller with the original figure.
That came from reading the statutes, not the brief. Let a buyer improve
after seeing the field and the whole process becomes an auction, which needs a licensed
auctioneer in all three states.
03It caught itself leaking a counterparty
Bid requests mask the seller until they authorise release. A test asserted that. It
failed, because a lot’s internal title defaults to the seller’s own
company name, and that title was going straight into the subject line.
Buyer-facing text now builds its own title and scrubs seller name, contact,
email and city from every line, because plant names hide in spreadsheet location columns.
04It prices the wait to be paid
Offers are ranked on what the seller nets after freight, loading, inspection
allowance and fee, then after discounting the payment delay at 18% a year.
A $28,900 offer at 45 days should not automatically beat $27,200 at 21. The
seller is the one who waits, so the comparison prices the waiting.
05It refuses to be paid by both sides in silence
Taking a fee from the seller and the buyer on one transaction, undisclosed, risks
forfeiting the entire commission. Opening a deal that pays from both sides is blocked
until both disclosure dates and both consents are on record.
The safest structure is to be paid by one side only. The software makes
that the path of least resistance instead of something you remember on a good day.
06A headline number can never become a price
The source spreadsheet held nine market figures, such as “Fisher control valves
$15,000+”. Every one is stored with its limitation attached and is
structurally unable to enter a calculation.
Those are buyer marketing. Quote one to a seller as what they will receive
and you have made a promise you cannot keep, on the first call.
Demo two: the part that shows you
An agent that works unsupervised is only worth having if you can see what it did. That is the
real objection, and a dashboard is the answer to it. Not a nice-to-have bolted on at the end,
but the thing you open every morning to decide whether you still trust it.
So ours show the refusals, not just the wins. The panel below is styled entirely from
six CSS variables. Change the theme, or pick your own accent colour, and watch it become
somebody else’s product. That is the whole white-label argument, testable rather than
asserted.
Four themes and a colour picker, all changing the same six variables. A real
one carries your logo, your typeface and your language, and reads from whatever the agent is
actually running. The numbers above are sample data, as any dashboard demo is.
Three defects, found before anything shipped
Left in view deliberately. Software that never had a bug was never really tested.
- The self-checks were wiping the live desk. They shared a state directory
with the running system. Tests now run in a throwaway directory, so destroying real data to
run a test is structurally impossible.
- Seed data was handed out by reference. Confirming a buyer mutated the
shipped dataset in memory, so a later reset wrote the edit back out as though it were
original.
- The risk reserve scaled with the prize instead of the risk. Set as a
percentage of deal value, it blocked every healthy small deal. Now sized from actual
logistics exposure.
203 self-checks, and they test policy rather than plumbing. Not
“does this function return a number” but: does an unconfirmed buyer stay
unreachable, does a milestone refuse to advance without evidence, does a bid request leak a
name. Those are the checks that still matter in a year.
What this means if it is your business
The domain here happens to be industrial surplus. That part is incidental. The transferable
claim is this: you have documents that describe how your business runs, and nobody has
built the thing that runs it. A process lives across a slide deck, a spreadsheet, three
email templates and one person’s memory. The gap between those and working software is
the entire job.
| You already have | What it becomes |
|---|
| An SOP or a training deck | Stages that will not advance without their evidence |
| A pricing spreadsheet | A calculator that cannot show profit before costs |
| A call script | Briefs built from the live record, with the do-not list attached |
| A compliance checklist | Gates that stop the work, naming what would clear them |
| Contract templates in Word | Documents that fill themselves from the record |
| A list nobody maintains | A registry with expiry dates that refuses stale entries |
| A weekly status meeting | A dashboard that already answered it |
When this is the wrong purchase. If your process is genuinely one person
using judgment on every case, there is nothing to enforce and you should not buy it. If the
documents contradict each other, that has to be settled by you first, because software
makes disagreement permanent rather than resolving it. And if nobody follows the process today, a
system that enforces it will be resented rather than used.
What you can buy
An agent kit. Design the agent that would run a process like this one for your business,
and download everything it needs: its instructions, the tools it may use and what each one may
change, a workflow per capability with its gates, the records it keeps, the rules that stop it
acting beyond its authority, and the acceptance tests that say whether it works.
| Kit | For an agent that | Price |
|---|
| Operator kit |
owns one job, start to finish. |
$79 |
| Deputy kit |
runs a whole function across more systems and channels, with judgment and escalation. |
$199 |
The size of what you design decides which kit it is, and the kit page tells you why. It is a
one-off purchase, not a subscription. A kit is a specification you build with the tools you
already use, not software that runs by itself, and it says so on every page it comes from.
How buying works
The tools on this site are free and the prices are printed because you should be able to buy
this the way you buy anything else that works: without a sales process.
01Design it
Pick the job, the systems it works with, what it may do on its own, and the rules it
must never break.
02See every file
The kit page shows the full file list and the price before you decide, not in a
proposal a week later.
03Buy
Payment is on Gumroad. Your licence key is on the receipt.
04Download and build
Paste the key, download the kit, and build it with what you already use. Change the
design and download it again, up to 20 times.
There is no discovery call, no proposal, and no sales process. You never have
to speak to anybody.
Related
To see the diagnosis before the build: Business Scan reads your live website and
works out which jobs are costing you most. Agent Designer takes a description of how
your business runs and designs the agent for it. App Builder writes single-file
tools from a description. All three are free and run in your browser.
The catalog lists every role and capability;
the kit page designs the agent and prices its kit without a call. If you want the argument behind
all of this: what an agent actually is and
how one is kept under control.