Skip to content

eamonnmag/ai-agents-dpi

Repository files navigation

AI Agents & the Data Processing Inequality

→ Open the interactive visualisation

A principle from information theory that doesn't get enough attention in AI agent design.

The Data Processing Inequality states that information cannot increase through processing — any transformation can only preserve it or reduce it. Applied to chained AI agents, this means accuracy cannot improve as you add steps. Each agent is a lossy channel, and errors compound.

This tool makes that concrete. It shows four decay scenarios across a chain of up to ten agents — independent errors, correlated errors, best-case RAG-augmented pipelines, and worst-case raw recall — alongside real benchmark data from the AA-Omniscience evaluation (Artificial Analysis, Nov 2025).

What you can explore

  • Adjust base model accuracy and chain length to see how quickly accuracy degrades
  • Compare independent vs. correlated error regimes and understand why the difference matters in production
  • See where frontier models actually sit on the accuracy/hallucination spectrum
  • Understand why human review at pipeline checkpoints is an architectural requirement, not a workaround

Built with

React · TypeScript · Recharts · Tufte CSS

Running locally

npm install
npm run dev

About

Showing how data processing inequality impacts the reliability of agents

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors