# AI Agent Evaluation

The missing piece in today's agent architectures.

Rippletide evaluates agent outputs and outcomes, not prompts, at runtime, **before the answer reaches the user.**

## The problem

### Not the LLM kind of hallucination. The agent kind.

Rippletide's first module is focused on a single recurring problem in every agent stack: **hallucinations**. But not the LLM kind, the agent kind, the ones compounding in each multi-step process, including false claims about your products or documentation.

#### False product claims

Your agent invents features, prices, or specs that don't exist in your documentation.

#### Compounding errors

Each step in a multi-step process amplifies inaccuracies from the previous step.

#### Invisible failures

Bad answers look perfectly fluent. Without verification, you can't tell truth from fabrication.

## How it works

### Claim-level verification, at runtime.

For each candidate answer your agent prepares, Rippletide runs a three-step verification pipeline, before the response is delivered.

01

### Extract factual claims

For each candidate answer your agent prepares, Rippletide extracts every factual claim: entity, attribute, and relationship.

02

### Search your trusted data

Each claim is checked against an exhaustive hypergraph containing your trusted data, including your RAG index if you want.

03

### Verify or block

Every claim is classified as supported, unsupported, or contradicted. Rippletide sends back the information to block the answer and computes a hallucination rate.

## Why runtime matters

### Three approaches. Only one prevents damage.

**Pre-deployment testing**  
When: Before deployment  
What: Prompt evaluation  
Enforcement: Static  
Result: Tells you after the fact that something went wrong

**Output monitoring**  
When: After execution  
What: Log analysis  
Enforcement: Reactive  
Result: Alerts you after the damage is done

**Rippletide**  
When: At runtime, before delivery  
What: Claim-level verification  
Enforcement: Deterministic  
Result: Blocks hallucinations before the user sees them

### Outputs, not prompts

We evaluate what the agent actually says, not the instructions it received. Because the gap between intent and execution is where hallucinations live.

### Runtime, not post-hoc

Other approaches tell you after the fact that something went wrong. Rippletide performs evaluation during execution, **before a bad answer is returned.**

### Your data, your truth

We import everything you share, product docs, knowledge bases, your RAG index. Verification happens against **your** ground truth, not a generic model.

## Developer tools

### Eval Platform

Measure hallucinations, inspect detailed traces, and get production-ready guardrails, all from a single dashboard.

### CLI Tool

Evaluate AI agent endpoints directly from your terminal. Test against predefined questions, validate responses, get instant feedback with real-time progress tracking.

### SDK

Integrate Rippletide evaluation directly into your agent pipeline with our SDK. Framework-agnostic, drop-in integration.

## Runtime Agent Evaluation

### Stop shipping hallucinations.

Rippletide is the runtime evaluation layer that catches agent hallucinations before they reach your users. Less than 1% hallucination rate. 100% guardrail compliance. Full explainability.