Skip to content
VZU
VZU OS · Agent

Agent 10 of 12

ORACLE.

Production prompt engineer.

Oracle is the VZU Prompts agent. Production system prompt and agent blueprint capability. Versioned prompts, eval suites, deploy config. The agent that ships prompts that work in production.

Brief VZU on ORACLE → Open prompts.vzu.ca →

The role

What ORACLE does

ORACLE is the dedicated agent behind VZU Prompts. It runs every time the operator briefs the vzu prompts capability. It owns the goal, the toolset, the memory, the audit trail.

The goal

What ORACLE is trying to achieve

"Write, version, test, and deploy production system prompts for AI agents. Run eval suites. Iterate."

The capability

What it powers

ORACLE powers the VZU Prompts capability in the VZU family.

Visit prompts.vzu.ca →

The toolset

4 MCP actions ORACLE can call.

Every action is exposed as an MCP server (Model Context Protocol). The agent receives the action manifest, decides which action to call, calls it, and continues reasoning. The audit trail records every call.

  • mcp.prompt.write

    Write the system prompt

  • mcp.eval.run

    Run the eval suite

  • mcp.prompt.iterate

    Iterate based on eval results

  • mcp.deploy.staging

    Deploy to staging

The memory schema

What ORACLE remembers between calls.

  • → The prompt versions
  • → The eval results
  • → The deploy config

The brief

A real demo brief.

"Build the production prompt for the voice agent. Inbound, outbound, escalation, handoff. 50+ eval cases."

Below is ORACLE's actual execution trace for this brief — every step, every MCP call, every observation, every audit trail entry.

The execution trace

What ORACLE did.

oracle@vzu-os — trace.log
  1. 01 ◉ oracle > read brief
  2. 02 ◐ oracle > thinking: voice agent prompt. 4 intents. 50+ eval cases.
  3. 03 ◇ oracle > mcp.prompt.write > v1: 1,247 tokens, 4 intents
  4. 04 ◇ oracle > mcp.eval.run > 47/50 pass, 3 fail (escalation edge cases)
  5. 05 ◇ oracle > mcp.prompt.iterate > v2: add explicit escalation policy
  6. 06 ◇ oracle > mcp.eval.run > 50/50 pass
  7. 07 ◇ oracle > mcp.deploy.staging > staged to staging.atlas.vzu.ca
  8. 08 ⏵ oracle > ship. 1 week. 50 eval cases. 1,247 token prompt.
oracle>

Audit trail · 6 entries

  • T+00:00:00 brief.received
  • T+00:00:30 prompt.v1.drafted
  • T+00:02:00 eval.v1.run
  • T+00:04:00 prompt.v2.iterated
  • T+00:06:00 eval.v2.passed
  • T+01:00:00 brief.completed

The numbers

ORACLE, by the numbers.

  • 1 week

    Median prompt duration

  • 50

    Eval cases

  • 100%

    Eval pass rate

  • 1,200

    Avg token count

The other 11

Every agent in the family.

Get in touch

Brief ORACLE.

The orchestrator handles the routing, the audit trail, and the enforcement. ORACLE handles the work. The operator owns the goal.

Brief VZU → prompts.vzu.ca →