LLM Output Validation
Zod validation with retry and error feedback, JSON extraction from messy model output, and reusable schema pieces: helpers a host wraps around its own LLM calls.
Evaluation Guide
Write test cases, score an agent's answers with built-in or custom scorers, grade with an LLM judge, compare runs and print reports.
Prompt Caching
AgentOS caches LLM prompt prefixes by default on every provider that supports
Cache Diagnostics
Anthropic prompt caching only pays off while the beginning of a request stays
Cost Optimization
What an AgentOS application spends is set by which model answers, how many tokens each call sends and receives, and how many calls a turn makes. This page lists the controls for each, and how to measure spend before changing anything.
Uncensored Content & Policy-Tier Routing
A policyTier of 'mature' or 'private-adult' routes a call to models that accept adult content. What it changes depends on the API: