Decentralized Artificial Intelligence through Controlled Emergence (DICE)

The DICE program seeks to advance the state of artificial intelligence by developing decentralized coordination and local inference control mechanisms to enable scalable, adaptive, and resilient collectives of heterogeneous AI agents capable of autonomously executing long-duration missions in contested environments while remaining under human oversight. Unlike centralized orchestration or ad hoc agent compositions, DICE leverages principles of self-organization akin to the internet’s robust infrastructure, where predictable, aligned global behavior emerges from simple local rules. The program is structured across three technical areas: TA1 focuses on peer-to-peer coordination through distributed auctioning, multi-agent reinforcement learning, and fault-tolerant consensus methods; TA2 develops control tools such as role-drift detectors, mission-alignment metrics, and inference-time controllers to ensure agent behavior stays within intended parameters; and TA3 builds a publicly accessible testing and evaluation platform, simulation environments, and metrics to assess scalability, adaptability, and resilience, including a state-of-the-art centralized baseline for comparison. Deliverables include algorithm implementations, courseware, simulation frameworks, and continuous reporting through quarterly updates and biweekly DARPA meetings, with all systems required to be demonstrated through a series of phased evaluations. The solicitation HR001126S0010, issued by DARPA under the Department of Defense, has a submission deadline of August 25, 2026, and requires proposers to submit both a technical and management volume (limited to 25 pages) and a detailed cost proposal using DARPA’s standardized spreadsheet. Proposals must conform to strict mandatory technical requirements and possess a valid and current CMMC Level 1 or 2 certification, depending on whether the work involves Federal Contract Information or Controlled Unclassified Information, with CMMC requirements flowing down to all subcontractors handling such data. Work is anticipated to be conducted at unclassified levels for TA1 and TA2, while TA3 will handle CUI, requiring compliance with DFARS 252.204-7012, NIST 800-171, and DoDI 5200.48 for storage, processing, and reporting of security incidents. Key personnel must commit substantial time to the effort, and proposers must disclose organizational conflicts of interest, provide SAM.gov UEI, CAGE code, and TIN, and certify their small business status if applicable. The full program spans 36 months, divided into three phases with specific milestones including annual workshops in Washington D.C., Boston, and San Francisco,