DARPA’s DICE Program

The Defense Advanced Research Projects Agency (DARPA), through its Information Innovation Office (I2O), has released a Special Notice titled Decentralized Artificial Intelligence through Controlled Emergence (DICE), DARPA-SN-26-65. This notice reflects DARPA’s ongoing mandate to pursue bold, high-impact research that delivers asymmetric advantages for national defense. With DICE, DARPA is targeting one of the most pressing challenges in modern AI development: enabling large, heterogeneous collections of AI agents to coordinate autonomously in contested, dynamic environments while remaining reliably under human control. The overarching goal is to move beyond brittle, centralized AI orchestration toward a decentralized architecture where robust collective behavior emerges predictably from local rules, mirroring the resilience of the internet itself.

Future conflicts, as DARPA envisions them, will unfold at machine speed across highly dynamic and contested battlespaces. Meeting this challenge will require autonomous, multi-agent AI systems capable of sustained, long-time-horizon missions without continuous centralized direction. The DICE program, led by I2O Program Manager Susmit Jha, seeks to develop the foundational theory and algorithms to make this possible.

DICE aims to create a decentralized AI architecture in which agents dynamically form teams through peer-to-peer coordination, execute complex missions autonomously, and remain resilient to agent failure, compromise, or adversarial interference, including the risk of “rogue” agents developing misaligned instrumental goals. The program draws on two complementary bodies of research: the theory of self-organizing systems and distributed consensus algorithms, and recent breakthroughs in controlling the internal reasoning of AI foundation models at inference-time (inference-time compute control).

The program’s technical approach centers on two core capabilities:

• Decentralized Coordination: Developing algorithms that allow AI agents to self-organize into teams, adapt to evolving mission parameters, and maintain collective coherence without a central controller, ensuring scalability and resilience even as individual agents are lost or compromised.
• Local Inference Control: Implementing mechanisms on individual agents that enforce role coherence, suppress misbehaviors, and constrain emergent collective behavior to remain aligned with commander’s intent across multiple inference steps and extended mission durations.

DICE is explicitly focused on the algorithmic and theoretical foundations of decentralized AI. The program will use simulation environments to demonstrate its architectures in Department of Defense-relevant use cases, with measurable targets for scalability, adaptability, and resilience against both benign failures and adversarial attacks. Development and deployment of autonomous systems in the real world is outside the program’s scope.