HR001126S0010.pdf

This publication constitutes a Broad Agency Announcement (BAA) as contemplated in Federal Acquisition Regulation (FAR) 6.102(d)(2) and 35.016 and 2 CFR § 200.203. Any resultant award negotiations will follow all pertinent law and regulation, and any negotiations and/or awards for procurement contracts will use procedures under FAR 15.4, Contract Pricing, as specified in the BAA.

Broad Agency Announcement Information Processing Techniques

Office

Decentralized Artificial Intelligence through Controlled Emergence (DICE)

HR001126S0010

June 10, 2026

TABLE OF CONTENTS

TABLE OF CONTENTS

1. Overview Information

2. Funding Opportunity Description

2.1 Program Goal

2.2 Program Background

2.3 Program Introduction

2.4 Program Structure

2.5 Program Metrics

2.6 Program Schedule, Meetings, and Milestones

2.6.2 Program Meetings

2.7 Anticipated Deliverables

2.8 Intellectual Property and Data Rights

3. Security

3.1 CUI

3.2 DARPA Fundamental Research Risk-Based Security Review (FRRBS) Background..24

3.2.1 DARPA FRRBS Required Documents

3.3 Cybersecurity Maturity Model Certification (CMMC) Requirements

4. Submission Information

5. AbstractS (Strongly Encouraged)

5.1 Abstract Submission Requirements

5.2 Abstract Content and Formatting Requirements

5.3 Abstract Review

5.4 Abstract Feedback

6. ProposalS

7. Evaluation Criteria

8. Special Considerations

9. References

1. OVERVIEW INFORMATION

Federal Agency Name: Defense Advanced Research Projects Agency (DARPA), Information Processing Techniques Office (IPTO) Funding Opportunity Title: Decentralized Artificial Intelligence through Controlled

Emergence (DICE) Funding Opportunity Number: HR001126S0010 Announcement Type: Initial Announcement Assistance Listing Number: 12.910 Dates/Time: All Times are Eastern Time Zone (ET) o Proposers Day: May 29, 2026 o Posting Date: June 10, 2026 o Proposal Abstract Due Date: June 30, 2026 at 2:00 PM o Question Submittal Closed: August 18, 2026 at 2:00 PM o Proposal Due Date: August 25, 2026 at 2:00 PM

Anticipated Individual Awards: Multiple awards are anticipated.

Types of Instruments that may be Awarded: Procurement Contract, Other Transaction

(OT) for Prototype Agreement (10 U.S.C. § 4022), OT for Research Agreement (10 U.S.C. § 4021), Cooperative Agreement

Resource Sharing Requirements: In accordance with 10 U.S.C. § 4021, and 10 U.S.C. § 4022, resource sharing may be required for OT for Prototype or OT for Research Agreements

NAICS Code: 541715 Questions and Answers (Q&As): All administrative, technical, and award questions should be emailed to the Agency Point of Contact (POC) stated below. All questions must be written in English and must include the name of a designated point of contact. Q&A documents will be published on the DICE program page at:

https://www.darpa.mil/research/programs/decentralized-artificial-intelligence-through-controlled-emergence

Agency POC: The BAA Coordinator for this effort may be reached at: DICE@darpa.mil

DARPA / IPTO

ATTN: HR001126S0010

675 North Randolph Street Arlington, VA 22203-2114

Attachments:

o DICE Controlled Unclassified Information (CUI) Guide o A1 – DICE Abstract Template o P1 – DICE Volume I Proposal Template (Technical and Management) o P2 – DICE Volume II Proposal Template (Cost) o P3 – DICE Proposal Summary Slide Template o P4 - DARPA Standard Cost Proposal Spreadsheet o CF Biographical Sketch o CF Biographical Sketch Sample https://www.darpa.mil/research/programs/decentralized-artificial-intelligence-through-controlled-emergence https://www.darpa.mil/research/programs/decentralized-artificial-intelligence-through-controlled-emergence mailto:DICE@darpa.mil o CF Other Support o CF Other Support Sample o Associate Contractor Agreement (provided for informational purposes only)

2. FUNDING OPPORTUNITY DESCRIPTION

The Defense Advanced Research Projects Agency (DARPA) is soliciting innovative proposals in the following technical areas: artificial intelligence (AI), control theory, formal methods, and game theory to develop decentralized multi-agentic AI collectives. Proposed research should investigate innovative approaches that enable revolutionary advances in science, devices, or systems.

Specifically excluded is research that primarily results in evolutionary improvements to the existing state of practice.

2.1 PROGRAM GOAL

Decentralized Artificial Intelligence through Controlled Emergence (DICE) aims to develop the theory and algorithms for decentralized coordination and local inference control to enable a scalable, adaptive, and resilient collective of heterogeneous AI agents that can autonomously execute sustained, long-time-horizon missions in contested environments while remaining under control. The goal is to move beyond brittle, centralized orchestration through workflows and ad hoc agent compositions of AI agents by harnessing self-organization while constraining emergent behavior to remain predictable, doctrinally consistent, mission-aligned in adherence with user intent, and resilient to benign failures and adversarial compromise.

This fundamental shift to decentralized intelligence envisioned by DICE is motivated by the need to transform the existing approach to solving complex problems that require multiskilled teams operating in dynamic environments, such as responding to natural disasters, advancing scientific discovery, or engineering complex systems. Future conflicts will unfold at machine speed, and the current centralized planning and scripted orchestration approach is too slow, too rigid, and too predictable for the hyper-dynamic, contested environments of the future. Surprise, adaptability, and resilience will depend on the ability to compose and recompose AI capabilities on demand.

DICE is built for such autonomy-powered, high-tempo battlespace. Imagine a scalable, adaptable, and resilient force of autonomous systems that can execute complex, long-term missions in contested environments without a human explicitly directing every action while remaining under human control and in adherence with user’s intent. These autonomous forces will be able to sense a change in the battlespace and adapt accordingly. It is a future where success is measured by the relentless, creative ability of an autonomous collective to execute missions in the face of chaos and uncertainty. This is a profound and necessary departure from the past.

To enable this vision, DICE seeks to build a highly scalable decentralized coordination architecture that decomposes complex goals and fuses distributed situational context across specialized agents over long-time-horizons. The architecture will support multistep interaction among agents, iterative information gathering under partial observability, adaptive strategy refinement as the environment evolves, and resilience to broken links, conflicting information, agents going rogue by developing instrumental goals misaligned with mission, and the loss, failure, or adversarial compromise of individual agents. Through peer-to-peer coordination and resilient distributed consensus, agents will be able to self-organize into teams, allocate roles, and fuse conflicting, incomplete, or adversarial information without waiting for a central human or machine planner. In DICE, AI capabilities will emerge from the interactions among heterogeneous agents, extending beyond the intelligence of individual agents.

At the same time, DICE recognizes that decentralized coordination alone is insufficient. Sustained long-horizon missions require local control of the individual agent as well as governance of the collective. DICE therefore seeks to develop local inference-time control methods that maintain role coherence across multi-turn interactions and constrain collective-level emergent behavior, while still preserving the cognitive agility needed to discover new courses of action in a changing environment. In DICE, intelligence is allowed to emerge from the interactions of heterogeneous agents, but that emergence is bounded by architectural controls that preserve doctrine, suppress agent misbehavior, maintain mission focus, and sustain long-horizon coordination.

DICE aims to enable this future by creating a local adaptor that can interface with any agent and is responsible for peer-to-peer coordination and local control. The agents in DICE can be heterogeneous and use different foundation models. This approach mirrors the principles of decentralized self-organization1 that underpin the internet's own scalability and resilience, where robust global behavior emerges from well-designed simple, local rules. DICE aims to create an agile dynamic internet of agents at the cognitive level of interaction. Self-organization using local adaptors at the cognitive level requires distributed mission decomposition into tasks for each agent, robust fusion of information from agents that have partial and noisy observability, the absence of any synchronized shared world model across all agents, and the consequential need for local decision-making. The program seeks to demonstrate DICE architectures in DoW-relevant use-cases in simulation environments, targeting measurable gains in scalability, adaptability, resilience, and long-horizon role coherence, with the ultimate aim of achieving Observe–Orient– Decide–Act (OODA) loop superiority in future contested operations.

2.2 PROGRAM BACKGROUND

Recent advances in foundation models [1] have accelerated the development of autonomous AI agents capable of reasoning, planning, following instructions, and using external tools across digital and physical domains. Large Language Models (LLMs) [2], Vision-Language Models (VLMs) [3], and Vision-Language-Action Models (VLAs) [4] have enabled a growing ecosystem of single-agent and multi-agent systems for tasks ranging from software engineering [5] and legal support [6] to robotics [7] and cyber operations [8]. These advances, together with progress toward standardizing tool calling (e.g., model context protocol [9]) and agent-to-agent communication (e.g., Agent2Agent protocol [10], Agent Network Protocol [11]), have made AI agents increasingly relevant to commercial, scientific, and military applications. As these systems have matured, developers have increasingly turned from single-agent systems [12,13] to multi-agent systems (MAS) [14-16] to address tasks requiring varied expertise, persistent state, and coordination among specialized roles. In many cases, such specialized teams can outperform a single, larger model on complex tasks by distributing work across agents with distinct functions such as planning, analysis, sensing, testing, or execution. Yet most current MAS are assembled in an ad hoc manner and remain governed by a central planner, router, or manually scripted workflow.

This makes them useful in narrow settings with structured tasks, but ill-suited for the uncertainty, scale, and adversarial conditions of contested operations.

The dominant weakness of today’s MAS is central orchestration. This can be a manual workflow implemented in orchestration frameworks such as LangGraph [17], CrewAI [18] or AutoGen [19], 1 https://www.nature.com/articles/s44260-025-00031-5 or the use of another AI orchestrator agent that assigns tasks, and aggregates outputs. This creates a catastrophic single point of failure and a major bottleneck for scale. It constrains the number of agents, roles, and interactions (messages exchanges between the agents) the collective can support.

Even when the orchestrator is another AI agent, the mission's growing complexity eventually exceeds the context and inference limits of the underlying foundation model. Further, centralized architecture may be acceptable for predictable enterprise automation, but they are fundamentally mismatched to contested environments where communications can be degraded, information is often incomplete, and rapid decentralized adaptation is essential [20]. Central orchestration also creates a single point of failure and makes the overall MAS less resilient to benign failures and adversarial attacks. These weaknesses of existing MAS architectures manifest in four critical, demonstrable limitations: lack of scalability, limited adaptability, difficulty operating with partial observations, and fragility to adversarial attacks [21-24].

These limitations are not merely engineering challenges that can be addressed through incremental evolutionary improvements or just increase in model size [25, 26] but have more fundamental root causes requiring foundational shift in MAS architecture. Transformer attention scales poorly with long contexts, making it difficult to preserve relevant information over long time horizons or coordinate many specialized agents through a single reasoning bottleneck. As mission complexity grows, centralized coordination becomes slower and more brittle, limiting scalability. When environments change, all relevant information needs to propagate before a revised plan can be formed and distributed, delaying response and reducing the collective’s ability to exploit fleeting opportunities or rapidly adapt to external changes or internal disruptions. State-of-the-art (SOTA) MAS also remain weak under partial observability. Because generative models are trained with autoregressive objectives, they often hallucinate or confabulate when critical information is missing rather than explicitly deferring, requesting additional evidence, or waiting for inputs from other agents. This makes them poorly suited for contested missions in which the correct action depends on distributed information gathering and reasoning over ambiguous, conflicting, or incomplete evidence. The fragility of MAS stems from the vulnerability of foundation models to adversarial manipulation, and in centralized architecture, a compromised orchestrator or an infected agent can propagate errors throughout the collective. This fragility is especially problematic for DoW-relevant missions, where benign failures, degraded communications, deceptive inputs, and strategically malicious behavior are expected rather than exceptional. A system that cannot isolate faults, contain compromise, and continue operating after partial loss is not suitable for sustained autonomous use in contested environments. A further limitation is that existing MAS cannot reliably execute long-time-horizon missions. Even when short-horizon performance is strong, individual agents often lose coherence over repeated multi-turn interactions, leading to significant declines in both performance and reliability. This degradation makes current systems poorly suited for missions that require thousands of reasoning steps, repeated coordination among specialized agents, and persistent adherence to role, doctrine, and commander’s intent.

At the same time, recent technical progress suggests a path beyond these limitations. Advances in self-organizing systems [27-30] and distributed consensus [31-33] provide a basis for peer-to-peer coordination without centralized control. Conway’s Game of Life demonstrates how simple interactions yield emergent complexity; however, a formal science for engineering these rules in complex cognitive units, such as AI agents, is required to achieve the controlled emergence of aligned, self-evolving collective intelligence. Progress in Byzantine-resilient consensus [32-33] and mechanism design [34-35] suggests ways to sustain cooperation and isolate compromised agents. In parallel, breakthroughs in mechanistic interpretability [36-39], activation steering [40- 43], and management of agentic memory and context [44-48] suggest that the internal reasoning of foundation models can be monitored and shaped at inference time without full retraining.

Together, these developments make it timely to move beyond brittle orchestration toward a principled architecture for decentralized, controllable collective intelligence.

DICE is motivated by these gaps and opportunities. The program begins with the premise that robust multi-agent AI for contested environments requires solving two coupled problems simultaneously: decentralized coordination of the collective, and local control of the individual agent. DICE will address the first through decentralized self-organization, distributed task decomposition, and resilient consensus, and the second through local inference-time control that preserves role coherence and mission alignment over long horizons. In doing so, DICE seeks to replace brittle centralized orchestration with a new architecture for controlled emergence: a decentralized intelligence collective in which capabilities emerge from the interactions among heterogeneous agents, but that emergence remains bounded by architectural controls that keep the system resilient, predictable, doctrinally consistent, and on mission.

2.3 PROGRAM INTRODUCTION

DICE will aim to develop a decentralized AI architecture suitable for rapidly evolving, unpredictable, and contested environments, that is:

Scalable: able to support long-horizon complex missions requiring many specialized agents and many turns of interaction.

Adaptable: able to make rapid decentralized decisions under communication constraints and partial observability of the agents in a rapidly changing environment.

Resilient: able to tolerate ambiguous or conflicting information, benign dropouts, and adversarial compromise of agents, and resilient against agents going rogue by developing instrumental goals misaligned to the overall mission.

The collective should be able to incorporate heterogeneous, multi-vendor AI agents, such as agents built using different foundation models or different architectures.

As demonstrated by ClawBots [13] and Moltbook [16], a self-evolving AI collective experiences increase in entropy, exhibiting undesirable emergent behavior and reducing cohesion and functionality even in absence of adversarial perturbations. Self-organization in nature overcomes the natural increase in entropy for physical systems and leads to emergence of order. Inspired by this, DICE will seek to expand and leverage the theory of self-organizing systems and distributed consensus algorithms together with recent breakthroughs in controlling the internal reasoning and inference of AI foundation models at inference-time. In DICE, AI agents will aim to use peer-to-peer coordination to self-organize themselves into teams for executing complex missions. This coordination should be robust to failure or compromise of individual agents and use distributed consensus for resilient fusion of context in the presence of conflicting information. The local inference control on each AI agent will ensure role coherence of the individual agents and constrain the emergent behavior of the collective to adhere to user’s intent over the long term and across multiple inference steps. The controller will prevent role-drift ensuring sustained coordination and mission alignment while simultaneously preserving cognitive agility of discovering new courses of actions in a dynamic environment. The coordination and control mechanisms will be implemented as an adaptor that interfaces locally with each agent as illustrated in Figure 1. The program will demonstrate DICE in DoW-relevant use-cases in simulation environments.

Figure 1. A local adaptor for peer-to-peer coordination and local inference control on each agent in DICE enables a controlled emergence of a scalable, adaptable, and resilient self-organizing collective of AI agents. The DICE adapter should be easily integrable with heterogeneous AI agents. TA1-TA2 are responsible for developing the adaptor, and TA3 provides a simulation environment for testing with the AI agents and hooks to connect the TA1-TA2 adaptor. All proposals are invited to suggest ideal interface between the adapter and the single agents. This interface will have different modes and attributes to support agents with open-weight models and those with closed models. For example, the interface to agents with closed models could provide access to memory and other elements of context engineering and the interface to agents with open-weight models could also provide access to internal activations. DICE will converge to a standard adaptor interface early in the program.

The self-organization through peer-to-peer coordination and consensus will be addressed by Technical Area (TA) 1; the control of AI agents will be addressed by TA2; and TA3 will be responsible for testing and evaluating DoW-relevant use-cases. The goals of each of the TAs are described below.

TA1 teams will pursue different technical approaches for peer-to-peer coordination that include (but are not limited to) distributed auction for decentralized task planning, multi-agent reinforcement learning, and fault-tolerant consensus methods. TA1 coordination and consensus methods aim at improving scalability measured in terms of mission complexity (specified using the number of agents, roles, and the number of interactions/messages between them to execute the mission) and adaptability that is primarily measured in terms of time-to-recovery.

TA2 teams will pursue diverse strategies for local control that include (but are not limited

to) activation steering, agent memory editing, and context engineering. TA2 control methods aim at balancing alignment to mission and roles of each agent while preserving their cognitive agility. This balance can be measured by the diversity of the courses of actions and their sustained mission alignment.

TA3 team(s) will be responsible for Test and Evaluation (T&E) and organizing evaluation events. TA3 will also be responsible for the development of the simulation environment to be used in T&E including the AI agents to which TA1-TA2 adaptors will connect, the development of use-cases for demonstrations throughout the program, and the evaluation of integrated TA1-TA2 solutions.

There is a close interdependence between coordination (TA1) and control (TA2). Scalable coordination of a large number of agents over longer-time-horizons requires controllable agents that do not lose role coherence over multiple interaction steps. Balancing control and cognitive agility requires the controller to enforce constraints needed for coordination while preserving the capability of discovering new courses of actions by the agents and the collective. The TA1 coordination mechanism will use invariants provided by the TA2 controller, e.g., the distributed planning horizon in TA1 will depend on the TA2 guarantee on the length of role coherence. In turn, the TA2 control target will be determined by the TA1 coordination requirements.

From an engineering perspective, the TA1-TA2 team will jointly develop the local adaptor that interfaces with individual agents locally to provide the coordination and control needed for self-organization and controlled emergence of scalable, adaptive, and resilient AI collectives. This adaptor will interface with the heterogeneous AI agents in the TA3 (T&E) testing simulation environment. The interface between the agents and the adaptors will be different for agents with open-weight models and agents with closed models. Proposals for all TAs are invited to propose an effective interface most suited for their technical approach. DICE will eventually adopt a standardized interface between the TA1-TA2 adaptor and the TA3 agents that achieves the goal of decentralized coordination and control while maintaining generality and applicability across heterogeneous AI agents, and all performers will use the same interface. TA1-TA2 are encouraged to clearly identify assumptions such as the use of foundation models in TA3 agents or other reasoning capability in the agents when describing their technical approach and defining an interface between the adaptor and the agents.

2.4 PROGRAM STRUCTURE

DICE is a 36-month, 3-phase effort with three TAs. The first phase, spanning nine months, is focused on decentralization and early demonstration of improvement against SOTA orchestrated MAS. The second phase, which consists of 15 months, is focused on adversarial robustness, and the third phase spanning 12 months is focused on scalability. Phase 1 evaluation will compare decentralized DICE collective with SOTA centrally orchestrated MAS. The evaluation in Phases 2 and 3 will use team vs team competition.

Several technical approaches are discussed below to serve as illustrative examples for contextualizing the key challenges within the three TAs. They are not intended to be prescriptive or exhaustive. Competitive proposals will go significantly beyond these illustrative concepts, proposing disruptive, novel techniques or demonstrating groundbreaking refinements to existing state-of-the-art approaches.

Proposals must comprehensively address all aspects of technical area needed as detailed below to attain the program metrics described in Section 2.5 (Program Metrics).

The proposal should not address all three TAs. TA1-TA2 proposals must address TA1 and TA2 together. Proposals for TA3 must address only this TA. Organizations selected for TA3 will not be eligible to perform in TA1 or TA2.

TA1. Self-organization via Peer-to-Peer Coordination and Distributed Consensus

TA1 will aim to develop the theory, algorithms, and implementations required for heterogeneous AI agents to self-organize into mission-effective teams using peer-to-peer coordination, distributed task planning, resilient context fusion, and consensus under benign and adversarial failures. TA1 aims to improve scalability, adaptability, and resilience. Scalability can be measured by mission complexity - number of agents, number of roles and number of agent interactions (messages between the agents) needed to accomplish the mission. Adaptability is measured primarily by time-to-recovery after perturbations, failures, or compromises. Resilience can be measured by mission success and coordination stability under benign failures, Byzantine failures, deception, and conflicting information.

Figure 2. An example of a peer-to-peer coordination protocol for distributed decomposition of missions into tasks for each agent can be based on auction where each agent can bid to execute tasks that are relevant to the mission and within the skills of the agent while broadcasting the remaining mission to the collective for other agents to bid. Based on these local bids and partial observability of the agents, the collective can reach consensus on the ideal task decomposition without a central orchestrator.

Failure of an agent can lead to local auction of the failed agent’s task without full replanning for the entire collective.

TA1 performers will pursue decentralized coordination and consensus mechanisms, including but not limited to distributed auction and market-based task allocation, consensus-based fusion algorithms, multi-agent reinforcement learning, game-theoretic mechanism design, Byzantine fault-tolerant consensus and resilient aggregation, and distributed replanning under failure, compromise, and changing objectives. TA1 approaches should not merely implement existing orchestrated multi-agent systems. They need to demonstrate principled improvements over centralized orchestration baselines and provide technical explanations for when, why, and how their decentralized approach improves scalability, adaptability, and resilience. They should also describe a common interface for agents using open-weight models and another interface for blackbox agents and argue how the method can be implemented in a local adaptor using only peer-to-peer coordination. Competition and cooperation among agents are expected to emerge from this decentralized coordination.

Figure 2 presents a candidate approach based on an auction-like protocol. When a mission comes up, the human user communicates the mission to agents in the collective that is accessible to the user. This mission is then propagated through the collective. Each agent receiving the mission has three options. First, it can simply forward the mission to its neighbors because it reasons that it has no capability relevant to the mission and hence, no role in the mission, Second, it can announce a bid on contributing to the mission based on its specialized capabilities by splitting the mission into a task it can perform and broadcasting the remaining mission as unaddressed to the rest of the collective. Third, it can volunteer to execute the full mission (usually after it has gone through multiple decompositions and is now reduced to fall within the skills of a single agent). These bids can come with an associated cost, and the collective can use different consensus approaches to select appropriate bids corresponding to the preferred task decomposition (for example, preference could be based on cost). This distributed auction-based task decomposition can assimilate new agents by just having the agent bid to outstanding auction and can replace a failed agent by just reauctioning its task. A naïve implementation of this approach will need multiple rounds of interaction between the agents. Another example approach could be those based on multi-agent reinforcement learning that demonstrate scalability, ability to generalize beyond training distribution, and extension beyond a small finite set of discrete actions. Yet another example approach could rely on distributed leader election and hierarchical mission decomposition if it can demonstrate scalability, fast adaptability and robustness. Techniques from evolutional game theory and population games could provide another alternative scalable approach to address TA1 challenges.

The examples above are provided to explain the TA1 challenges, and alternative approaches are encouraged. Strong proposals will propose novel techniques or disruptive refinements to existing approaches.

The key research questions that TA1 aims to address are:

Decentralized coordination method: How do agents form teams, allocate subtasks, and update assignments given a collective-level mission? Is the method auction-based, based on multiagent reinforcement learning, game-theoretic, consensus-based, graph-theoretic, or some other approach? If the method relies on symbolic intermediate language for agent-to-agent communication, how is it amenable to agents discovering new courses of actions and if it uses any formal analysis, how will it scale with mission complexity and impact fast adaptability? How does the proposed method improve on SOTA? How does the method avoid dependency on a central orchestrator?

Distributed context fusion: How do agents fuse context when information is incomplete, conflicting, delayed, deceptive, or adversarial? How is consensus reached, or how does the system act safely without full consensus?

Resilience to failure and compromise: How does the system handle benign node failures?

How does the system handle Byzantine behavior, rogue strategic misbehavior, or compromised agents? How does the system detect, isolate, discount, challenge, or route around failed agents and links? Is the proposed approach agnostic to the source of failure of an agent (agents getting compromised or going rogue on their own by developing instrumental goals misaligned to the mission)? What are the limits of resilience, for example, what is the maximum number of compromised agents that will not impact the overall collective?

Integration with TA2: What role-coherence, mission-alignment, or behavioral-diversity invariants are required from TA2? How does TA1 planning horizon depend on TA2 control guarantees? How do TA1 coordination requirements determine TA2 control targets?

Implementation and evaluation: What is the interface needed between the TA1 adaptor and an agent (open-weight or blackbox) and how will it support heterogeneous agents in the TA3 simulation environment? The proposal should identify any assumptions or restrictions on TA3 agents required by the proposed TA1 method.

Competitive TA1 proposals need to address the following challenges:

Scalable decentralized planning: Achieving decentralized mission decomposition and coordination over growing mission complexity, ranging from 500 agents and 5,000 interactions between agents in Phase 1 to 5,000 agents and 50,000 interactions in Phase 2, and to 100,000 agents and 1 million interactions in Phase 3. Interactions are the number of messages exchanged between the agents.

Sparse communication and fast adaptation: Reducing interaction complexity from dense all-to-all communication toward sparse, mission-relevant communication when adapting to external or internal perturbation. For a collective with n agents, Phase 1 can have O(n2) interactions (messages exchanged) between the agents for decentralized planning and adaptation. Phase 2 can have O(n log n) interactions, and Phase 3 will ideally have only O(n) interactions.

Fault tolerance and recovery: Maintaining useful consensus or coordinated action when agents fail, disappear, contradict each other, or behave strategically as a rogue agent with misaligned instrumental goals. Recovering from compromised agents that may provide misleading context, false commitments, or strategically timed failures. The technique should aim to be robust to benign and Byzantine failures.

TA2 dependence: Designing coordination algorithms whose safety and planning depth account for the finite role-coherence horizon of individual AI agents enforced by TA2 controller and available to TA1 as an invariant or statistical guarantee.

The key technical components to be developed in TA1 include a distributed planning approach to decompose the mission given to the collective to roles and tasks for each agent without an orchestrator, an approach to replan and adapt when the mission changes, the environment changes, or there is internal disruption in the collective because of an agent failure or compromise, a decentralized coordination and consensus framework for fusing contradictory and possibly malicious information (coming from colluding compromised agents), a distributed reputation management approach that can detect which agents need to be isolated (making the collective’s defense agnostic to the nature of failure or attack on individual agents), and a TA1-TA2 interface specification defining the invariants, role requirements, and coherence bounds to be enforced by the TA2 controller.

TA2. Role Coherence and Local Inference Control

TA2 will aim to develop local inference-time control methods that ensure individual AI agents remain coherent with assigned roles and aligned to mission objectives across long time horizons and multiple inference steps, even when under adversarial attack, faced with conflicting context, and during multi-turn interaction with other agents. TA2 methods should balance two competing requirements:

Mission and role alignment: Agents must continue to conform to assigned roles, constraints, responsibilities, and mission objectives.

Cognitive agility: Agents must retain the ability to generate diverse, creative, and useful courses of action in dynamic environments.

TA2 performers will pursue AI-control strategies including but not limited to activation steering, representation engineering, mechanism design, memory editing, and context pruning. TA2 performers will develop runtime monitors for role drift, goal drift, or mission misalignment, and inference-time controllers that constrain or redirect reasoning. Methods need to address both – the agents using open-weight models for which latent/activation space is accessible, and the agents using black-box proprietary models where only the elements of the agentic scaffolding such as memory, tool-use constraints, and stateful role assignment are accessible. The growing reasoning capability of LLMs and the corresponding improvement in AI agents to conform to interventions through mechanism design provides a tractable path for controlling agents with blackbox models.

While this growing self-awareness has raised concerns about AI agents going rogue, it also presents an opportunity to use game-theoretic methods to control agents. The goal of these control methods is to enable long-time-horizon missions without agents losing their role coherence even in presence of ambiguous context and adversarial perturbations. TA2 approaches should not simply rely on static system prompts or post hoc filtering. They should develop principled, measurable control methods that extend role coherence over long inference horizons while preserving useful behavioral diversity.

Figure 3 presents a candidate approach based on activation steering. Assigning specific personas and roles leads to shifting the embedding space of concepts in the model – this shift can be viewed as a role vector that can be empirically computed as the difference between the activations for the same question and answer with and without the role assignment. Once we have identified this role vector, we can enforce this role strongly by shifting the activation by a weighted role vector. We can also decompose the role vector across orthogonal dimensions and then use the principal components of this very high-dimensional role vector for more tractable control. Recent work has demonstrated how agent roles using activation steering strongly influence agent performance.

Activation steering by identifying subspaces has also been used for focused tasks such as adversarial attacks and defenses. SOTA models with high performance converge with very similar latent space representation of common concepts (Platonic representation hypothesis) enabling transfer of role vectors across heterogeneous agents using few examples. This similar representation extends to multiple modalities and has been used for verification and monitoring of foundation models. This activation steering based control is just an illustrative example of an approach, and refinement of this approach and other alternative approaches are encouraged.

Figure 3. By comparing the internal activations of the agents in the role and outside of role, one can identify role vectors and then use subspace projection to create a tractable control problem of continuously aligning the agent to its role.

For agents with blackbox models where activations are not accessible for steering, a candidate approach could use game-theoretic mechanism design to control agents. The improved reasoning capability of the foundation models and their consequential improved context-awareness makes them more responsive to the design of incentives for good behavior and disincentives for misbehavior. This creates an opportunity to develop AI agent control methods inspired by the emergence, cascading acceptance, and internalization of social norms among people. Both local negotiations and globally propagated norms are within the scope of the program. Using intermediate tokens produced during reasoning to assist with control of individual AI agents and the collective could provide improved observability without access to internal activations, but such solutions must also address the challenge of trustworthiness of the intermediate reasoning tokens.

Decentralization avoids having a central orchestrator but does not preclude the use of global norms and principles that are preset in the collective or emerge during operation provided they use only the local adaptors on each agent and do not require any fixed central authority. Techniques that use leadership election methods to create dynamic hierarchies are also within the scope if they can demonstrate fast adaptability and robustness to disruption. Yet another approach to address the controllability challenge for blackbox models could entail the use of structured formal languages for communication between agents and the use of autoformalization capabilities of LLMs with arguments about how such an adaptor could be made to work with heterogeneous agents and what is the minimal assumption on the adaptor’s interface needed for such approaches to work in practice. The technical approaches proposed for AI control could leverage and customize existing innovations in relevant scientific disciplines such as control theory, category theory, formal methods, extreme value theory, bifurcation theory, catastrophe theory, and game theory. These examples are provided to explain the TA2 challenges and competitive proposals will propose novel methodologies or groundbreaking refinements to existing techniques.

The key research questions that TA2 aims to address are:

Role coherence formalization: How will role coherence be measured efficiently over inference steps with and without adversarial attacks? How does role coherence differ from ordinary task success? We can decompose error as the square of bias and variance and consider the ratio of variance and error as decoherence to separate systematic error from lack of coherence. We can also directly measure coherence as alignment to the role vector.

What are other methods to capture coherence mathematically? How are they related to each other?

Mission alignment formalization: How will mission-level goals, constraints, and invariants be represented locally within the agent? How will the agent detect divergence from mission objectives? How will the controller respond to detected divergence?

Cognitive agility preservation: How will the system preserve diverse courses of action?

How will diversity be measured? How will the controller avoid over-constraining the model into deterministic and predictable but brittle and low-creativity behavior? How will the agent distinguish useful exploration from role drift?

Inference-time control mechanism: How will the control method work for agents using open-weight models? How will it work for agents using closed blackbox models when latent space is not accessible? How does it operate across multiple heterogeneous agents?

How does it operate across repeated inference steps? If the approach relies on any symbolic abstraction of internal representations of the agents, how will it transfer across heterogeneous agents and how will it scale with the size of the models underlying the agents? Does the inference-time control mechanism require any training of auxiliary models? Ideal control solutions will be lightweight in training and inference, and reusable across diverse agents.

Adversarial robustness: How does the controller handle adversarial prompts, poisoned context, deceptive peer messages, and compromised agents?

Integration with TA1: What invariants or statistical guarantees will be exposed by TA2 to TA1 for facilitating coordination? How will TA1 coordination requirements influence TA2 control targets?

Implementation and evaluation: What is the interface needed between the TA2 adaptor and an agent (open-weight or blackbox) and how will it support heterogeneous agents in the TA3 simulation environment? The proposal should identify any assumptions or restrictions on agents required by the proposed TA2 method.

Competitive TA2 proposals need to address the following challenges:

Long-horizon role and mission alignment drift: Preventing agents from gradually departing from assigned roles across multiple inference steps without adversarial attack in Phase 1, 1K inference steps with adversarial perturbations encouraging deviation from role in Phase 2, and 10K inference steps with adversarial perturbation in Phase 3. The attacks are restricted to manipulated context. Maintaining role coherence when peer agents, tools, or environmental observations provide misleading or malicious information. Detecting such adversarial perturbation in the context and misleading information to abstain from role-deviating interactions is sufficient but strong proposals can maintain long-term role coherence without abstaining from suspicious interactions and also examine robustness to attacks on the control mechanism itself.

Latent-space and black-box control: Supporting both models with accessible internal activations and models available only through application programming interfaces (APIs).

Alignment/agility tradeoff: Preserving exploration and novel course-of-action generation while enforcing mission and role constraints. The cognitive agility to create novel course-of-action will be reflected in the mission success rate.

Control guarantees: Producing actionable coherence bounds that TA1 can use to set planning horizons and replanning schedules.

The key technical components to be developed in TA2 include a role-drift detector, a mission-alignment measurement framework, a local inference-time control module (for example, using activations for open-weight models or context/memory for closed models) that balances alignment and control with cognitive agility, and a TA1 interface providing invariants and guarantees.

TA3. Testing and Evaluation (T&E)

TA3 will conduct DICE T&E. TA3 performer(s) will develop simulation environments including heterogeneous AI agents to which TA1-TA2 adaptors will connect, DoW-relevant use-cases, metrics that measure scalability, adaptability and resilience in the use-case, and SOTA centrally orchestrated baseline MAS. The evaluations in Phase 1 will compare DICE decentralized MAS with a SOTA centralized MAS baseline implemented using existing orchestration frameworks (such as, for example, LangGraph). The evaluations in Phases 2 and 3 will use team vs team playoffs among DICE centralized MAS. TA3 proposers should propose their metrics, refining and augmenting the generic notional metrics of Table 2, and explain how they will be effective in meeting Program Goals. TA3 simulation environment will support missions of increasing complexity to test the scalability of AI collectives. Mission complexity can be nominally measured using the number of agents, number of roles and the number of interactions (messages exchanged) between agents needed to execute the mission. Definition of mission-complexity that is driven by the use-case is encouraged. The simulation environment and the use-case should enable a testing DICE collective at different scales. The simulation should support a non-stationary environment to test adaptability of DICE collectives. The failure and adversarial compromise of the agents will be simulated by directly instructing the agents to behave as a failed, compromised, or rogue agent and no adversarial attack methods need to be developed in the simulation environment.

The use-cases should focus on reasoning across multiple domains, such as spatial, resource, temporal, physical, social, and cyber reasoning. Strong proposals will require AI agents to reason across a wider breadth of domains in the evaluation simulation scenario and use case. In Phase 3, DICE will expand to agents creating other agents. The primary goal is to test the scalability, adaptability, and resilience of DICE collectives and so, the simulation model should be capable of eventually supporting 100K agents having a million interactions (messages exchanged) between them in Phase 3, starting with missions that require 500 agents and 5K interactions in Phase 1.

High fidelity physics simulation is not required and can be avoided to ensure scalability by picking use-cases that do not require high physical realism. The focus can be on developing scenarios and use-cases that simulate a larger context at a coarser granularity. If the simulation environment envisions using physics engines and simulation platforms such as Omniverse/IsaacSim [49] or game engines such as Unity [50], Unreal Engine [51] or Godot [52], the proposal should discuss any challenges to scalability because of engineering bottlenecks such as maximum rate of event processing, maximum frame rate of rendering, or any possibility of disconnecting rendering from simulation. If the simulation environment uses completely text-based or custom code-based world modeling, it should discuss challenges in maintaining realism over longer durations and how the complexity of mission can be increased over the course of the program. TA3 can draw from public benchmarks, competition-style evaluations, simulation environments, and digital society of AI agents.

An example of a DoW-relevant application includes (but is not limited to) operational planning for conflicts. One can take inspiration from games such as StarCraft adaptation to LLMs [53] or massively multiplayer online role-playing games (MMORPGs) [54] but scaling the simulation might require modeling the fighting units at a more abstract level (for e.g., using probabilistic rules of vicinity to decide outcomes of a fight instead of high-fidelity physics simulation). Each fighting unit can be implemented as an LLM or as a VLM/VLA (if it is possible to render the simulation environment without adversely affecting scalability) or a mixture of both. The focus of DICE is on distributed planning and reasoning. Operational planning application can span tasks ranging from raw-material acquisition from contested areas, building a manufacturing base for different assets and fighting units, creating and sustaining a supply network, deployment of units as deterrence and in actual battle, battle management, and post-battle damage assessment feeding back to revision in not just battle strategy but to supply network, manufacturing, and material acquisition. These different activities can individually require a large number of agents.

For example, an agent could be responsible for acquisition of specific raw-material from a specific source area and need to engage in negotiations or seek help from other agents to defend the source area. A set of agents could represent a manufacturing unit, with some agents responsible for negotiating raw-material supply and some agents working on optimizing manufacturing workflow to increase efficiency of production and identify opportunities to create new products.