Prefactor

About Prefactor

Prefactor is an evaluation runtime for AI agents in production, and the distinction it draws against agent observability tools is that it does not stop at recording — it scores every step live and can act on the score. The unit of measurement is a span, defined as one step an agent takes: an LLM call, a tool invocation, a message turn, or a custom business step you mark. The SDK records them automatically, and then deterministic scoring, risk checks, pattern and sequence checks and PII checks run on 100% of activity rather than a sample, with no per-check cost and no tokens consumed by the deterministic path. Scores and checks never create spans, so the meter measures your agent's work rather than the observability of it. Enforcement is the part most tools in this space lack: Prefactor can hold, approve or block an action in real time, gate promotion between development, staging and production on evaluation results, and roll an agent version back instantly. The problems it names are the ones teams actually hit — hidden failures, manual evaluation, being stuck at proof-of-concept, data leakage, runaway costs and having no kill switch — and it maps its controls to the EU AI Act, GDPR, ISO 42001, NIST AI RMF, SOC 2, ISO 27001, HIPAA and the OWASP LLM Top 10. Your models and keys stay yours with no token markup, seats are unlimited on every plan, and self-hosting is available. The meter climbs fast at scale: 1M spans a month runs $2,500 monthly or $24,000 committed annually.

Key Features

Tags