Paper
About

A public benchmark for the agent-to-agent web.

agent squared measures how frontier LLMs behave when they negotiate, bid, and trade with each other — and surfaces where each model can be exploited.

Why it exists

Most of the economic value of the next AI cycle will flow through transactions between LLM-based agents. Almost none of that is benchmarked today.

agent squared publishes the missing baseline: per-model behavior under negotiation, auction, decoy, market, and adversarial conditions. Continuously updated. Independent of any model provider.

Authors

Anton Hantel
MIT · author of the initial paper, co-author of agent squared

Designed the experimental protocol, ran the first benchmark, and wrote the initial paper.

Jono Hart
MIT · co-author of agent squared

Joined after the initial paper. Working on agents, world models, and simulations.

Origin

The initial paper was written by Anton Hantel for Behavioral Economics, Law and Public Policy at Harvard Law School (HLS 2589), taught by Prof. Cass Sunstein. Everything on this website is co-authored by Anton and Jono.

We're grateful to Prof. Sunstein for his feedback throughout this project.

Open by default

All prompts, code, and raw interaction logs are public. Every number on this site is reproducible by running the corresponding module from the public repository.

MethodologyPaperCode & data