Introduction

SmartSnap is a major breakthrough in the search for solutions for autonomous agents. The technology, developed by Meta, focuses on creating agents that can proactively and scalably verify their own performance.

How SmartSnap Works

SmartSnap uses an innovative approach to enable autonomous agents to verify their performance. The idea is that the agent creates a series of evidence snapshots, which are then evaluated by a large language model (LLM) to determine their validity and relevance.

Principles of Self-Verifying Agent

The Self-Verifying Agent is designed with two primary missions: completing a task and verifying its own performance through curated evidence snapshots. The agent uses its online access to perform verification in situ, without requiring external verification from a human.

Experiments and Results

Experiments conducted on mobile tasks across various model families have demonstrated that SmartSnap enables LLM-driven agents to be trained in a scalable manner, achieving improvements of 26.08% and 16.66% respectively for large models.