Implicit Intelligence and Agent‑as‑a-World: Evaluating agents on what users don’t say

Implicit Intelligence is a framework designed to evaluate an agent's ability to understand and act upon under-specified real-world requests by inferring missing constraints, promoting reasoning and context awareness rather than merely following explicit instructions. This approach utilizes Agent-as-a-World (AaW), which allows environments to be defined in a single natural language file, enabling models to simulate realistic scenarios without complex coding. By testing agents with scenarios that mimic real-life tasks, such as muting a phone during an appointment or turning off lights while considering ongoing activities, Implicit Intelligence assesses an agent's capability to perceive the environment, interpret feedback, plan strategically, and act effectively without detailed guidance. This framework emphasizes the importance of agents operating safely and intelligently in unpredictable environments by focusing on emergent reasoning, adaptability, and ethical decision-making, ultimately measuring not just the actions agents take but their underlying motivations and ability to navigate nuanced scenarios.