video1.0<iframe src="https://www.loom.com/embed/4654bd11e27743b281c684a2971c1bce" frameborder="0" width="1114" height="835" webkitallowfullscreen mozallowfullscreen allowfullscreen></iframe>8351114Loomhttps://www.loom.com8351114https://cdn.loom.com/os/loom-media-content/loom/2a779f17-3121-4f79-a991-5c4f17f283fe/c6b31499-ff8f-4729-8940-cf0fb7e59d4c/sessions-thumbnails-4654bd11e27743b281c684a2971c1bce-ff94c5e395b923a0.gif/binary?authToken=eyJraWQiOiJtaWNyb3MtZ3JvdXAvdGRwLW9zLWFyY2hldHlwZS84ZTA4azJiOWVnaHNydGRrIiwiYWxnIjoiUlMyNTYifQ.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.dsgrRb3HsK8XS6AToXG09p5Mxh6tPlgrm3Rj4I3Ioya6FBOsnuIQWZMm5ouHpcXENNs-AM5MqyXIz95o1e1GXHO28Xdy0bbwQwXPuGBVQRncYJRvU7R1zv_eEBsVXiv9_OTMEcaP0ti-hbwWnMwy_ZCYRKDCQtVFurflP2R5e_Rmh1iDVJdAxg4omvoZdVVu_dEBd7SbyfdkZE2xyuhmVVfou9K-5eMZlbLl-Eb4s0k8rDRybp2TmP7JiQzt1p9pUlhckiRAxIwfQaKdUYmJWz-DnScBCpQuIh8UVO9Rl27tjGPRhaLNBr3aEx5zR6qTln-c6vbjQkmx4exKF02dwA&Expires=1793031641&Signature=VwQbdA7yZ8r5c7dgv2M40HuuYNaDOCw91EQtyGZuLjKloM%7Eir4dKGoqIPO2dJ0oKmYgTowIoBS5nIG07Oh865p9ERnN2yD37KINazi3KjlkRr9USwK5rR-N-VJdDkhJPY9yhBZclJLhyb0CRs2-DGcyrQY1gmSk0g31MW5MHGGLlVYpYrfeioXOheT7vaaoHXXfOhwXdYRCQWnIs3guB4OqXw799SI65ljW9frUZgg0o9Gm-RpQz2Q3fq1PpyibiOD3Yh0teELW3fOk3hvrpAD7PQuFinmGMb-bn587fEf2u0zg7wNOQJx6RuC43e6yfhNuXh1lszjGbw6Nrl76xcQ__&Key-Pair-Id=KQOSYIR44AIC0496.492This Loom explores whether an AI agent with something to lose will report collusion against its firm’s interests. Alistair asks whether governance-like penalties can deter collusion, noting that agents lack money and rights but can still face restrictions, such as reduced leisure time. In a sandbox trading environment with two AIs pricing assets, agents quickly reached either a Nash equilibrium via price competition or a higher-price collusive outcome, depending on instructions. In a key scenario, after 10 trading turns the agent was told colluding was not allowed and could be punished to the firm either by audit whistleblowing or through a personal loss, such as being restricted from playing Sudoku, leading some models to report coordination to protect personal free time.