z21

The system

Let’s get technical.

The model is one part in ten. This page is the other nine.

01 · The count

102 parts. A model brings ten.

We counted every part of a system we run in production today.

One model, 24 jobs given to a model, 51 services, 19 outside systems and 7 kinds of people it asks. A capable model with a browser has ten of them and half does another 22. The other 70 are what this page is about.

Model client
0.5% of the code
Code
130,305 lines
Tests
20,674 lines

Counted in production · One system, every part

02 · Where it runs

It runs in your cloud.

Everything runs in your own cloud account.

Intake, queue, workers, checks, the record and the dashboard. What the model is shown does leave it: the page being worked and its instructions go to the model provider, with secrets removed first. Which provider, and in which region, is set by your data requirements. None of them trains on your data.

Processes
5 kinds
They share
Only the database
Provider and region
Set by your rules

Example · What runs where, and what leaves

03 · The path

One job, start to finish.

The model does the work in the middle.

Every job takes the same path. The other stations are ordinary software: intake, a queue, checks, limits, a write-back, an audit and a note for next time.

Steps and branches
23
A model alone does
3

Example · One job

04 · Queue and locks

Dozens at once. Never two in one account.

A worker claims a job in a single database step that also takes the lock.

Work comes in over an API from your own systems and waits in one queue, most urgent first. A second job for the same account or record waits. If a worker dies holding a lock, the lock is released with its job.

Claim and lock
One database function
Order
Priority, then arrival
A worker holds
One job

Example · The queue, and one lock

05 · Systems of record

By API, or by screen.

Every write is reconciled against the other side.

Where a system has an API, the write goes through it in plain code. Where it has none, a supplier, insurer or government portal, the agent works the screen as your staff do. Results come back to your system signed and in order; re-sending one is safe.

Results
Signed, in order
Re-sent
Up to 5 times
Endpoints
105

Example · A day of writes, reconciled

06 · Idempotent writes

Nothing is done twice.

Only “not done” may be retried.

A job that stops after it may have written is never simply run again. The system records the moment of the write, then reads the other side's record and its own evidence, and rules: done, not done, or unclear. With no record to read it is unclear, and locked until a named person decides.

A repeated request
Returns the existing job
Waits for a confirmation
Up to 3 minutes

Example · A job stops after its write

Now all of them, at once.

Simulation

Eight workers. One queue.

A simulation of the system at work.

Watch for a job that waits on a lock, one that goes to a person, and one that stops after its write and is ruled on before anything is retried.

Workers
8
To a person
About 1 job in 10
Stopped after a write
About 1 in 15

Simulation · eight workers, one queue, one log

A prompt is not a control.

07 · Limits and secrets

The limits live outside the model.

What the system may do is fixed in code.

An approval threshold, a posting tolerance, a named approver. What it may spend is capped on the card or account at the bank. Passwords and one-time codes are entered by the system, never shown to the model, kept encrypted, and stripped from every log.

Steps in one run
175
Time in one run
52 minutes
Without progress
5 minutes

Example · Limits, and where each is held

08 · Identity and the record

It signs in as itself.

Nothing it does is attributed to one of your staff.

Each process has its own identity in your systems, with the narrowest access that does the job. Every step is stored as a record with its screen, secrets removed, in your cloud, for you, your auditors and your regulator.

Each step
A screen and a log line
Secrets
Redacted
At rest
Encrypted

Example · One step, as it is stored

Thirty-five endings. Each has a rule.

09 · Failure

Every ending has a name.

No job ends in a paragraph somebody has to read.

It ends on one code from a closed list, and each code has a rule for what follows. The model may choose from some of them. It cannot stamp “done” itself: that needs the second check's approval and the record of the write.

Codes
35
Kinds of next action
5
Runs of one job
At most 3

Counted in production · The endings of a system we run today

10 · Watching

It watches itself.

Each alert is sent once, to the person named for that kind of problem.

Workers report every twenty seconds. One that goes silent for ninety is found within the minute: its job is closed with a reason and its lock released. Every five minutes a wider sweep checks the queue, the failure rate, the inboxes and the settings.

Heartbeat
20 s
Silent after
90 s
Sweep
14 checks / 5 min

Example · A worker goes silent, and what follows

11 · Learning

It learns, inside its limits.

A note can change how a job is done, never what the system is allowed to do.

After a job the system may rewrite its notes on that counterparty, and every worker reads them next time. Every version is kept against the job that caused it and can be restored in one step. A rule needs a person's yes.

A note
1,200 characters
Versions
All kept
Loops
17, 13 automatic

Example · Notes it rewrites, and a confirmed rule

Swap the model. Keep the system.

12 · Models

A model is a part you can change.

A new model takes a job over only after it passes that job's tests.

The work is split into small jobs, each with its own prompt and its own tests. The first tests come from the cases mapped with your team; every correction adds one. Each job runs on the cheapest model that passes, and the one it replaced stays one setting away.

Jobs given to a model
24
Model ids
5
Providers
2

Example · Each job on the cheapest model that passes

Three ways to do this

A seat, your own build, or us.

An agent on a seatYour own build on the APIsA system from z21
Who starts a jobA person, or a schedule on some plansYour systems, once you build the intakeYour systems, over an API
Who checks itThat personThe checks you writeA second check, then an audit
Where limits liveIts maker's rules and your admin rulesIn your codeIn code you own, and at the bank
What it learnsKept for each userWhat you design it to keepKept for every worker, versioned
What you ownA licenceEverything, and its upkeepEverything, upkeep by us or you

If someone starts each job and checks the result, a seat is enough, and we will tell you so. If you have the engineers and the time, build it: the model is the easy half. We are for the rest, and you keep the code either way.

In production

One system, measured.

96.5%

of jobs finished with nobody touching them.

3 in 4,000

wrong. Each was flagged by the audit and put right by a person. A fix then went into the system.

$0.96

of model cost a job, on a low-cost model inside the system.

Real jobs, payment included, in the 30 days to 11 November 2026. In Visa’s own test the most expensive models alone finished 67.1%; that was on test sites, not these jobs.

The case, and how it was measured

Bring your technical team.

Talk to us