---
title: Introduction
description: What Prefactor is, how the platform is structured, and where to
  find what you need.
editUrl: true
head: []
template: doc
sidebar:
  hidden: false
  attrs: {}
pagefind: true
draft: false
---

Prefactor records agent activity, classifies data risk from your activity schemas, and stores quality evidence you (or your eval pipeline) attach. You instrument agents with the SDK; every run is an [instance](/platform/concepts/instance) made up of [spans](/platform/concepts/span), one per step. Prefactor scores each instance against a configurable [risk profile](/platform/concepts/risk-profile), and quality evaluations attach to the instance (Prefactor also writes quality spans as an audit side-effect). The [web app](/admin-ui) is where you inspect runs, configure risk profiles, and review scores and quality results.

## From pilot to production

**Pilot.** Instrument your first agent with the SDK. You get span-level activity records — LLM calls, tool invocations, and agent spans — on the [Agents page](/admin-ui/agents) and the agent's [Instances tab](/admin-ui/agent/instances). The default schema covers you at this stage; no custom schema or risk profile is required. The [Quickstart](/quickstart) walks through a demo end to end.

**Adding structure.** Tighten [activity schema](/platform/concepts/activity-schema) definitions where you need clearer contracts. Once those definitions deploy, the web app shows conformance for each span type and flags gaps. Assign a [risk profile](/platform/concepts/risk-profile) so Prefactor classifies each run from the span types that ran and what those types declare. Declare quality schemas and start attaching evaluations when you need scored evidence on runs — see [Quality and performance](/platform/quality-and-performance).

**Going to production.** Create a Production [environment](/platform/concepts/environment) alongside Development. Pin a version per environment with an [agent deployment](/platform/concepts/agent-deployment), and use a deployment-scoped [API token](/platform/concepts/api-token) for the runtime. Purpose labels (**Live**, **Eval**, **Smoke test**) keep evaluation traffic distinct from production.

**Evidence and audit.** Every instance, span, schema version, risk classification, and attached quality payload is available for review — subject to what your integration sent under its capture and sampling settings. See [The audit trail](/platform/audit-trail).

## Concepts and explainers

- [Data model](/platform/data-model) — how account, agent, instance, span, and schema fit together.
- [Risk](/platform/risk) — what Prefactor means by data risk.
- [Sensitive data](/platform/sensitive-data) — marking, redaction, and discard.
- [Quality and performance](/platform/quality-and-performance) — evaluations vs risk vs business performance.
- [Instrumentation strategy](/platform/instrumentation-strategy) — how to grow schemas over time.
- [Audit trail](/platform/audit-trail) — what the connected record contains.