Customize a production-grade agent with your business processes, glossary, and guardrails — then ship it on infrastructure you control, with the whole platform underneath it.
Off-the-shelf agents miss your naming, your rules, and your edge cases — and there's rarely a safe way to encode that knowledge without hand-holding every answer.
What this layer gives you.
Manifest-driven templates spin up an agent bound to your workspace and connected data.
Author persona, business processes, reasoning frameworks, glossary, and source priorities.
Business processes and guardrails compile into declarative policy rules, enforced deterministically at runtime.
Author, snapshot-test, and activate vetted queries before they ever reach users.
Agents at the summit
Composed on the whole stack beneath them.
One integrated stack, GPU to Agent. This layer, in context.
Agents don't improvise against your data. They bind to deterministic primitives your team authors — Queries, Processes, Rules, and Templates — versioned, policy-checked, and auditable.
Anatomy of a run
The seam
Above: the runtime flow. Below: the resource layer your team authors. The LLM picks which Process runs — never what it does, never what it calls.
LLM chooses which Process. Your team authored the rest.
Click any building block to open its deep-dive ↓
Deep-dive
01 / 04
01 · Building block
Vetted SQL, parameterized. Written once by your data team. Every agent binds to the same query.
A Query is a canonical answer path — the same SQL your analyst wrote once, versioned in the platform, invoked identically whether a human asks through the app or an agent calls in over MCP. No LLM writes the SQL that touches production data; it chooses which Query to run and fills the parameters.
Why
One source of truth. Same query, same answer, every time — cross-checked, tested, versioned forever.
Authored by
Data team.
What it looks like
-- Query: revenue_by_segment -- Owner: data-team@vantedge -- Params: :period (yyyy-mm), :region (iso-2) -- Version: 4 · Last verified: 2026-06-30 SELECT segment, SUM(net_revenue) AS revenue, COUNT(DISTINCT customer_id) AS accounts, ROUND(AVG(net_revenue), 2) AS avg_per_account FROM finance.revenue_ledger WHERE fiscal_period = :period AND region = :region GROUP BY segment ORDER BY revenue DESC;
The four together
| 01Queries | 02Processes | 03Rules | 04Templates | |
|---|---|---|---|---|
| WhatWhat it is | Vetted SQL, parameterized. | End-to-end SOP: steps, checks, halts. | Declarative guardrails, compiled to a policy engine. | Fixed output shape — memo, spreadsheet, dashboard. |
| WhoWho authors it | Data team. | Ops, with the workflow owner. | Security or compliance. | Whoever owns the deliverable. |
| RuntimeWhat the LLM does | Chooses one. Fills parameters. Never rewrites the SQL. | Invokes one. Follows the steps. Halts where told. | Cannot bypass. Denials are logged. | Rotates prose. Never rewrites the numbers. |