VZU Portfolio · prompt-engineer
Prompts as production code. Versioned, tested, evaluated, audited.
A prompt-engineering surface on the Netwit Agentic OS runtime. Versioned prompts, regression tests, MCP contracts, and an audit trail for every model call.
Prompt Engineer is a prompt-engineering surface for the operator who treats prompts as production code. Prompts are typed — input schema, output schema, tool calls, fallbacks — all in the same file, all in version control. The test set is generated with the prompt and committed alongside it. The eval harness runs on a schedule. Multi-model routing is on by default. The MCP contract is enforced. The audit trail is the prompt log. Prompt Engineer is a tool inside the VZU family, operated by Virtual Zero Unbound Inc.
What ships
Lead capability
02
Input schema, output schema, tool calls, fallbacks — all in the same file, all in version control. The contract is the contract. The prompt is production code.
03
The agent writes a test set, the test set is committed, the test set runs on every change. Catch the regression before it ships.
04
Every tool call is typed, every input is validated, every output is parsed, every call is logged. The MCP contract is the integration story.
05
Route the same prompt across two models, read the diff, ship the better one. The routing decision is recorded in the prompt log.
06
Every change, every test, every eval, every routing decision. The audit trail of the prompt is in the repo, not in a vendor dashboard.
The deep view
The point of Prompt Engineer is the typing. A prompt is not a string in a config file. A prompt is a typed artifact with an input schema, an output schema, tool calls, and fallbacks. The contract is the contract. The agent writes the prompt, the contract, and the test set at the same time. The diff is in the repo. The reviewer agent reads the diff before the prompt is allowed to ship.
The regression tests are the safety net. The agent writes a test set, the test set is committed, the test set runs on every change. The eval harness reads the result. Pairwise comparison, single-rating, golden set — the operator picks the harness, the agent runs it on a schedule. The weekly eval summary lands in the inbox. The reviewer agent flags the regressions before the prompt is allowed to ship.
The MCP contract is the integration story. Every tool call is typed, every input is validated, every output is parsed, every call is logged. The contract drift is caught by the reviewer agent before the prompt is allowed to call the tool. The multi-model routing is on by default — the operator can route the same prompt across two models, read the diff, ship the better one. The routing decision is recorded in the prompt log.
The audit trail is the prompt log in the repo. Every change, every test, every eval, every routing decision. Not in a vendor dashboard, not in a Slack channel, not in an email thread. The audit trail of the prompt is the prompt. The operator can read it, version it, branch it, hand it to another team. The prompt is a real artifact with a real history.
The process
The prompt is typed: input, output, tool calls, fallbacks. The agent writes the test set with the prompt. The contract is the contract. The prompt is production code from the first commit.
The test set runs on every change. The eval harness reads the result. The reviewer agent flags regressions. Multi-model routing is configured at the same step. The eval is on a schedule.
MCP server contracts are typed, every tool call is validated, every output is parsed, every call is logged. The contract drift is caught by the reviewer agent before ship.
The prompt lands in the runtime. The routing decision is recorded. The audit trail is the prompt log. The weekly eval summary lands in the inbox.
Where the work happens
The Prompt Engineer surface is operated out of three VZU offices. The eval harness runs overnight, the prompt log lands in the repo, the contract drift is caught before the prompt is allowed to call the tool.
In production
Regulated firm
With a humans-in-the-loop override on the eval threshold. The prompt log is the regulator's form. The reviewer agent caught a contract drift in week 4; the team fixed it before it shipped.
SaaS classification
Read the diff, ship the better one, with a weekly eval summary. The routing decision is in the prompt log. The reviewer agent catches the regressions. The override rate dropped 40% in the first quarter.
Enterprise agent
Audit trail in the repo. The reviewer agent flags a contract drift on day 14; the team fixes it before it ships. The eval harness runs overnight, the summary lands in the inbox.
Other VZU tools
On the runtime
Prompt Engineer is a configuration of the Netwit Agentic OS runtime. The runtime enforces RBAC at the agent level, ships the audit trail by default, and deploys in cloud, hybrid, or on-prem modes. The prompt log in the repo is the audit trail.
The VZU platform, in detail →Questions
Get in touch
Email the team. Reply in one business day. The pilot is priced, the rubric is configured, the ship date is real.