# Prefactor

> Prefactor evaluates every AI agent run in real time, scoring quality, drift, and risk in production, then enforces policies to hold, approve, or block failing agents automatically.

Prefactor is a real-time AI agent evaluation and enforcement platform built by Prefactor Pty Ltd, a Melbourne-based startup backed by Antler, Black Nova VC, and Func Ventures. It instruments production AI agents via lightweight TypeScript and Python SDKs, scores every run for quality, drift, and risk the moment it happens, and wires those evaluations into runtime enforcement — pausing, blocking, or escalating risky actions before they execute.

## What It Is

Prefactor sits at the agent runtime layer, providing what the company describes as a "reliability trifecta": observe every run as traces and spans, evaluate with LLM-as-judge and deterministic checks, and act by holding or blocking high-risk actions via the SDK or API. Unlike pure observability tools that surface dashboards after the fact, Prefactor enforces decisions in milliseconds at execution time — before a sensitive action completes.

## The Observe → Evaluate → Act Loop

The platform is organized around three stages, each backed by specific capabilities:

- **Observe**: Real-time tracing of every model call, tool call, and custom span; agent registry for fleet inventory; cost tracking attributed per agent; and an immutable audit trail for compliance evidence.
- **Evaluate**: Real-time quality scoring on 100% of activity using LLM-as-judge, technical checks, and qualitative metrics; composite risk scoring across data sensitivity and action consequence; PII detection across 17 sensitive-data categories.
- **Act**: Runtime policies that block, throttle, or escalate; human-in-the-loop approval routing that pauses a run and routes it to a reviewer with full context; PII deletion across every span the data touched; and a kill switch that stops an agent immediately without a code deploy.

Custom spans let teams pull context from any datasource — GitHub, Linear, Jira, internal databases, REST or GraphQL APIs — into the run, grounding every evaluation in what actually happened.

## Agent Lifecycle and Versioning

Prefactor versions every agent by semver, commit, or tag and tracks it through lifecycle states from pending to retired. Agents are promoted through dev, staging, and production only when their evals pass (eval-gated promotion), and instant rollback to a prior version is available from the dashboard. A two-tier schema-and-deployment versioning system surfaces conformance drift when a run breaks its declared shape.

## Framework and Stack Coverage

Native SDK integrations cover LangChain, Claude Agent SDK, CrewAI, Vercel AI SDK, OpenAI Agents SDK, Bedrock Agents, LlamaIndex, and more. The core TypeScript and Python SDKs plus OpenTelemetry ingest cover any other stack. Coding tools (VS Code, GitHub Copilot, Cursor, Claude Code) and workflow tools (n8n) are also listed as supported surfaces. The company states SDK overhead is under 1 ms per call.

## Enterprise Security and Compliance Posture

Prefactor is built around least-privilege access, full auditability, and existing identity stacks. Key security properties include AES-256 encryption at rest, TLS 1.3 in transit, per-environment API key scoping, PII detection on every span at no additional per-check cost, and an append-only immutable audit trail. The company states SOC 2 Type II is in progress, and the platform includes built-in risk mapping for GDPR, HIPAA, ISO 27001, NIST AI RMF, EU AI Act, and OWASP LLM Top 10. Data residency options cover Australia (primary), US, EU, and Asia Pacific, with enterprise self-hosted deployment available.

## Deployment and Audience

Prefactor targets platform, ML, and engineering teams at enterprises running agents in regulated industries — banking, legal, manufacturing, retail, and telecommunications are listed verticals. The company positions the product for developers, heads of AI, AI product managers, security and governance leads, and founders. The vendor states teams go from CLI install to first traced run in under five minutes, and instrumentation requires no migration or rip-and-replace of existing agent code. Runtime policies default to observation mode before enforcement is activated.

## Features
- Real-time agent run tracing and span recording
- LLM-as-judge quality scoring on 100% of production traffic
- Composite risk scoring across data sensitivity and action consequence
- PII detection across 17 sensitive-data categories
- Runtime enforcement: hold, approve, or block agent actions in milliseconds
- Human-in-the-loop approval routing with full run context
- Kill switch to stop an agent immediately without a code deploy
- Agent versioning by semver, commit, or tag with lifecycle states
- Eval-gated promotion through dev, staging, and production environments
- Instant rollback to prior agent versions
- Custom spans to pull context from any datasource into the run
- Cost tracking attributed per agent, tool call, and compute cycle
- Immutable append-only audit trail for compliance evidence
- Agent registry for full fleet inventory
- OpenTelemetry ingest for any stack
- TypeScript and Python core SDKs
- Native integrations for LangChain, Claude, CrewAI, Vercel AI SDK, OpenAI Agents SDK, Bedrock Agents, LlamaIndex
- PII deletion across every span the data touched
- Schema conformance drift detection
- Unlimited seats on every plan

## Integrations
LangChain, Claude Agent SDK, CrewAI, Vercel AI SDK, OpenAI Agents SDK, AWS Bedrock Agents, LlamaIndex, OpenClaw, LiveKit, n8n, VS Code, GitHub Copilot, Cursor, Claude Code, GitHub, Linear, Jira, PostgreSQL, OpenTelemetry, Sentry

## Platforms
WEB, API, VSC_EXTENSION, DEVELOPER_SDK, CLI

## Pricing
Freemium — Free tier available with paid upgrades

## Links
- Website: https://prefactor.tech
- Documentation: https://docs.prefactor.ai/
- EveryDev.ai: https://www.everydev.ai/tools/prefactor