AI Agent Guardrails
Guardrails are the checks that run on an agent's input and output so it doesn't leak data, get hijacked, or run away with your bill. agentFast runs them as a pipeline in the runtime — before the model ever sees the prompt, and again on everything that gets stored.
Obfuscation-hardened PII
Zero-width characters, homoglyph look-alikes and unusual spacing are normalized first, so obfuscated PII is still caught. Came out of an adversarial pass, with regression tests.
Injection neutralization
Instructions smuggled in tool output or retrieved content are flagged or neutralized rather than obeyed.
Abuse blocking
Unsafe requests are refused at the guardrail layer, not left to prompt-wrangling.
Rate limits
Token-bucket limits, global and per-tool, backed by Redis when you have it — so a runaway loop can't exhaust a downstream API.
Retrieval grounding
A confidence score on how well an answer is supported by retrieved context, so you can catch ungrounded claims.
On by config, every SDK
Declared in agentfast.yaml and enforced by the runtime, so they apply the same on LangGraph, CrewAI, or a vanilla loop.
Guardrails pair with observability (every intervention is a visible step) and human-in-the-loop (the last line before an irreversible action). See them on LangGraph or CrewAI.
agentFast is the production layer — memory, 50 tools, observability, evals, guardrails, human-in-the-loop — for LangGraph, CrewAI, the Claude & OpenAI Agent SDKs & Vanilla. Own it for life.
Get agentFast — from $89