Partly liveWhat works today is listed first. What is still to come is listed under Not yet.

VZN Agents

Approve what your agent proposes, and keep the receipt.

Agents that propose the work and wait for your APPROVE. The agent runtime: an agent loop, installable packs, a workflow engine and a Hub where every outbound action waits behind APPROVE. Live voice agents and personal assistants belong here too.

See it work

One job, one decision, one receipt.

Credit for a double charge, ticket #5120

job_01HX2
Example · scripted

Waiting for your APPROVE

Needs your approval

  1. Read the ticket
  2. Found the double charge
  3. Checked the credit policy

Credit R 120 to Naledi K. for a charge made twice on 1 November

Charge
R 120, twice
Past credits
0 in 12 months
Example success fee
R 35.00, only if it succeeds

Undo: A credit can be reversed until the invoice closes.

Approval is not proof the action ran. Turn this on to see a failed run.

Your receipt appears here once you decide.

It will carry the job id, who decided and when, the system's response, and whether anything is billed.

How it goes

Three steps.

  1. The agent proposesIt reads the ticket, drafts the action and shows its reasoning. Nothing has happened yet.
  2. You decideApprove or decline. Without your approval nothing moves and nothing is billed.
  3. A receipt is writtenThe job id, who decided and when, what the system answered, and what is billed, kept together.
Evidence

What we can show, and where it comes from.

What waits behind APPROVE
Every outbound action with consequencesSource: architecture/products/VZN-Labs-Agents.md
A receipt carries
Job id, decider, time, system response, evidence, feeSource: apps/web/src/landing/ApprovalDesk.tsx
A failed action
Keeps its evidence and bills nothingSource: apps/web/src/landing/ApprovalDesk.tsx
Example · scripted
job_01HX2 billing.credit · 200 · R 35.00 · receipt kept

Every action leaves a receipt: who approved it, when, and the proof that it happened. Nothing is billed without an approval.

What you get

Results, not a demo.

  • Every agent proposes work with its reasoning and waits for your APPROVE on anything with consequences.
  • Workflows are part of VZN Agents: multi-step jobs whose validators re-check customer text after substitution, so a customer's words cannot widen what a step does.
  • The hosted coding agent plans, codes and reviews on an agent branch; merge to pre-prod only when the reviewer and CI agree.
Proposed pricing

Pay for the tokens your agents use, not for seats that sit idle.

Self-hosted

R0 / month

Run the agent loop on your own model, your hardware, R0 per call.

  • Agent loop and workflow engine
  • APPROVE gate stays on
  • Self-hosted model in the registry

Pay as you go

Most common

Metered

Metered like the Hub: tokens in, tokens out, cache reads and batch discounts.

  • R5.55 in / R22.20 out per 1M tokens on DeepSeek V4.1 Flash
  • Per-job and per-day rand budgets
  • The meter's own ledger, nothing estimated

Enterprise

Talk to us

Hosted ops, private models and a support agreement.

  • Private deployment
  • Billing and compliance review
  • Talk to us

The usage units are the Hub meter's own: tokens_in, tokens_out, cache_read_tokens and zar_est.

Status

What runs today, and what is still open.

Works today

  • Agents Core pack, the agent loop and the workflow engine
  • Jobs, APPROVE and a results ledger in the Hub
  • A hosted coding agent (Agent Hub) that plans, codes and reviews on an agent branch

Not yet

  • Live voice agents
  • The in-platform coding orchestrator, which is specified but not built
Generative UI

Ask in words. Get a real screen back.

Instead of a wall of text, a VZN agent draws the screen: a rule, a table, a new agent. It is built from the same components as the rest of VZN, so you can edit it, and nothing happens until you approve.

Refunds over R2,000 should need a manager.

Here is the rule as it would run, and what it would have caught last month.

Scripted example. No model runs on this page and nothing is sent.

Drawing the screen…