Towards Sandboxes for the Internet of Agents
ADIA Lab Research Paper Series
Authors: Alessio Buscemi, Jordi Cabot, Pradyumna Chari, Francisco Herrera, Shadab Khan, Marcos López de Prado, Ramesh Raskar
Date Published: 28 January 2026
ADIA Lab's Advisory Board member Marcos López de Prado, and Health AI research lead Shadab Khan co-authored a roadmap paper on building evaluation infrastructure for the emerging Internet of Agents with collaborators at LIST, Univ. of Granada, and MIT.
As AI agents begin to discover and interact with each other through protocols like NANDA and A2A, a key question arises: how do you verify what an agent can safely and reliably do before letting it loose in a network?
The paper proposes a network of domain-specialized, interoperable sandboxes that evaluate agents across multiple dimensions - reasoning, tool use, bias, robustness, safety, and multi-agent coordination - and produce signed, replayable attestations that travel with the agent. These attestations are published as AgentFacts in the NANDA Index and verified at interaction time, closing the loop between evaluation and deployment.
Rather than scoring agents on isolated tasks, the approach measures system fitness: how populations of heterogeneous agents coordinate under realistic conditions including latency, partial information, and adversarial behavior.
Why this matters - as autonomous agents move from single-user tools to networked participants, trust infrastructure becomes critical. This work outlines a practical roadmap for certifiable, interoperable agent evaluation at scale.
