---
schema: 1
type: term
title: Reality gap
description: >-
  The **reality gap** is the distance between how an AI agent performs in a
  simulated testing environment (such as a simulated organization with synthetic
  data) and how it performs in real-world deployment. In this channel's
  material, Salesforce's Silvio Savarese describes using simulated organizations
  to stress-test agents and notes that closing this gap remains open work.
canonical: 'https://liquidgold.cc/channel/everyinc/explore/term/reality-gap-38c86a2b8ae0'
markdown: >-
  https://liquidgold.cc/channel/everyinc/explore/term/reality-gap-38c86a2b8ae0.md
llms: null
parent: 'https://liquidgold.cc/channel/everyinc/llms.txt'
channel: everyinc
updated: '2026-10-23T22:12:35.346Z'
language: null
children: []
---


# Reality gap

Canonical: https://liquidgold.cc/channel/everyinc/explore/term/reality-gap-38c86a2b8ae0

Parent: [Channel index](https://liquidgold.cc/channel/everyinc/llms.txt)

## Definition

The **reality gap** is the distance between how an AI agent performs in a simulated testing environment (such as a simulated organization with synthetic data) and how it performs in real-world deployment. In this channel's material, Salesforce's Silvio Savarese describes using simulated organizations to stress-test agents and notes that closing this gap remains open work.

## Reality gap in this channel's material The term appears in Dan Shipper's Dreamforce 2025 interview with **Silvio Savarese, Chief AI Scientist at Salesforce** [^50a3b3fa-9ad2-59a3-9c76-ddfe74635066@220]. In that conversation, Savarese explains how Salesforce stress-tests agents using **simulated organizations** with synthetic data — and he is explicit that closing the reality gap between those simulations and real-world deployment remains **open work** [^50a3b3fa-9ad2-59a3-9c76-ddfe74635066@676]. In other words, agents that perform well on synthetic, simulated benchmarks may still stumble when they face real enterprise data, real users, and real business conditions — and nobody in the material claims the problem is solved.  ## Using the term Use *reality gap* when discussing the risk that **evaluation in simulation overstates real-world performance**. It sits alongside the interview's broader diagnosis of why enterprise agents stumble — siloed data and jagged intelligence [^50a3b3fa-9ad2-59a3-9c76-ddfe74635066@431] [^50a3b3fa-9ad2-59a3-9c76-ddfe74635066@485] — but it is specifically about the simulation-to-production divide.  - The reality gap is the distance between how agents perform on simulated organizations with synthetic data and how they perform in real deployments. - Savarese names it as **still open work** — not a solved problem [^50a3b3fa-9ad2-59a3-9c76-ddfe74635066@676]. - It matters because enterprises stress-test agents in simulation before customers ever touch them, so a large gap means tests do not predict production behavior.