Prefactor

Prefactor is an observability and evaluation platform for AI agents already running in production. Its framing is that most tooling in this space reports problems after the fact — it can see everything and stop nothing — so Prefactor pairs tracing and scoring with an enforcement step that acts before an action is taken.

Every agent run is scored in real time for quality, drift and risk. Custom spans let teams enrich a run with context pulled from GitHub, a database or an internal API, so an evaluation is grounded in operational reality rather than the transcript alone. Where a run trips a rule, Prefactor can pause the risky action, route it to a human for approval, or block it outright — the enforcement loop that distinguishes it from a dashboard. It also detects 17 categories of sensitive information moving through agent traffic.

For teams shipping agent changes regularly, agent versioning supports a dev → staging → production path with eval-gated promotion, so a new prompt or tool definition has to clear its evaluations before it reaches users.

Native SDKs cover TypeScript and Python, with built-in support for LangChain, Claude, the Vercel AI SDK, OpenClaw and LiveKit, plus coding surfaces including VS Code, GitHub Copilot, Cursor and Claude Code, and workflow tooling such as n8n. OpenTelemetry compatibility covers anything else. The free tier includes 25,000 spans a month; paid plans start around $250 a month with usage billed near $2.00–$2.50 per 1,000 spans. Prefactor Pty Ltd is an Australian company backed by Antler, Black Nova VC and Func Ventures, with SOC 2 Type II and role-based access control listed on its roadmap.