DON27BZ02-NV012 TITLE: "Self-Replicating Evolutionary Agentic SI Swarming Force" with Autonomous Embedded SI Guardrail Enforcers

OUSW (R&E) CRITICAL TECHNOLOGY AREA(S): Applied Artificial Intelligence (AAI)

COMPONENT TECHNOLOGY PRIORITY AREA(S): Integrated Network Systems-of-Systems;Integrated Sensing and Cyber;Trusted AI and Autonomy

PROJECTED CMMC LEVEL REQUIREMENT: Level 2 (Self)

The technology within this topic is restricted under the International Traffic in Arms Regulation (ITAR), 22 CFR Parts 120-130, which controls the export and import of defense-related material and services, including export of sensitive technical data, or the Export Administration Regulation (EAR), 15 CFR Parts 730-774, which controls dual use items. Offerors must disclose any proposed use of foreign nationals (FNs), their country(ies) of origin, the type of visa or work permit possessed, and the statement of work (SOW) tasks intended for accomplishment by the FN(s) in accordance with the Announcement. Offerors are advised foreign nationals proposed to perform on this topic may be restricted due to the technical data under US Export Control Laws. 

OBJECTIVE: 1- Develop a platform agnostic fully autonomous Agentic SI swarm system suitable for multidomain naval operations (i.e., Sea, Air, Space, Land, and Cyber) with self-awareness for rapid operationally scalable response, namely "hyper-scalability". The Agentic SI system, a hybrid-evolutionary SI, must self-replicate and self-deploy. In other words, it must self-evolve as it gets feedback from its internal SI agents; it must coevolve as it interacts with the environment and the adversary to adapt and survive in littoral ecosystem without human guidance. As a team, it must autonomously reason to scale-up or scale-down its swarming clusters to generate tactical sub-clusters or end-game sub-clusters for defensive and offensive measures to confront and overwhelm threats. Neuro-Symbolic SI is to be a practical method for these SI-based swarm agents to explore and exploit the conditions in the operational ecosystem.

2- The developed platform must strictly adhere to operational policies, safety, and security with embedded SI agents serving as guardrails to control all sustaining SI agents (i.e., Self-Evolving SI agents) by checking outputs, cross-referencing data feeds and plans, and correcting flagged responses. These SI-enabled Guardrail-Enforcers must be active and adaptive systems embedded into the Agentic SI Swarm workflows. The embedded SI Guardrail-Enforcers validate-and-verify policies for additional safety and behavior guardrail checks beyond the ordinary trust metrics. The important performance question is "Did Agentic SI system get better without changing policy permissions, safety boundaries, or behavior unpredictably?"

DESCRIPTION: This SBIR topic seeks hardware-agnostic Agentic SI swarm-teaming technologies capable of self-replication, hybrid-evolution (self-evolution and coevolution), and self-organization into teaming many arrangements. It is important to note that this project does not challenge the importance of physical platforms/machines (i.e., unmanned assets) that have physical presence in operational environments. The envisioned Agentic SI system is the machine’s intelligent autonomy software that controls the platform’s presence and its interactions with entities in the operational environment. It autonomously configures individual swarms with unique missions and networks of collaborative swarms operating in multiple operational domains with specific teaming assignments.

The envisioned swarm system is agentic as it pursues goals autonomously. It interprets assigned mission objectives and determines how to achieve them. It knows how to process, fuse, and exploit all source intelligence (ASI) and live sensor feeds to determine the course of action (CoA) and keeps adjusting it until the outcome is reached. It decides which payloads to use and when to use them, for example, active and/or passive intelligence, surveillance, reconnaissance, and targeting (ISRT) sensors, point-to-point or omni-broadcast communication, or weapon systems, etc. It adapts when situations or environmental conditions change, or certain aspects of a plan fail. It can recover from various system failures and processing errors. It is designed with embedded SI-Guardrail-Enforcers to manage, validate, and verify every aspect of the Agentic SI system’s performance and associated risks to ensure that the end-to-end system adheres to established policies set forth by the Department of War and the U.S. Navy including operational safety requirements as well as the national security doctrines and mandates. The highest priority for this program is to establish ZERO tolerance for SI-generated rogue populations of unaligned agents rapidly and uncontrollably increase.

The hybrid-evolutionary aspect of the Agentic SI system focuses on attaining new operational proficiency and know-how to execute emerging missions with the highest accuracy, persistence, and endurance. The self-evolving SI system monitors endogenous performance of all subsystems, gets feedback, and improves its own internal design configuration autonomously, for example, altering or generating new internal models. The coevolutionary SI system monitors exogenous performance interactions and coevolves the internal design in response to interactions with the environment and adversaries. It improves autonomous teaming performance interactions and reconfigurations, for example, altering or generating new swarming models, plans, and configurations to maintain operational tempo.

The self-replication aims to create autonomous, self-improving agents capable of independent survival and expansion. This Agentic SI system autonomously improves itself, creates its own successors, and builds specialized sub-agents to solve operational issues.

The self-configuring aspect of the Agentic SI system focuses on a variety of mission-focused collaborative teaming arrangements, including scouting teams for early detection of indications and warning signs to ensure survival, trip-wire teams to detect encroaching threats, ad-hoc networking teams for continuous collaboration and tracking threats, a self-defense team for fault-tolerant operations, and offensive teams to ensure survival. In aggregate, these teams form an Agentic-Force that self-adapts, self-organizes, and self-heals.

The Agentic SI Swarming-Force addresses mission capability gaps that require agile adaptation to novel adversarial TTPs and unanticipated asymmetric countermeasures. The envisioned SI capability requires SI-based evolutionary modelling techniques for learning, reasoning, recognition, and autonomous decision-making to counter emergent and/or unanticipated observable patterns of life activities or events. In other words, agentic discovery of unknown-unknown events/assets posed by the intelligent adversaries. It ensures the timely execution of CoA plans to neutralize emerging threats, including asymmetric attacks, novel deployment of joint civilian and military assets, cloaked military assets, decoys, and deceptive asset deployments. A representative set of agentic impelling operations may be conducted overtly or covertly to influence hostile forces to undertake desired actions, including adopting specific courses of action or modifying their pattern of life to safeguard critical assets through reconfiguration, relocation, or repurposing. These observable responses can then be exploited by the Agentic SI Swarming Force to degrade adversarial autonomy and disrupt, reshape, and ultimately defeat their operational plans and tactical engagements.

A notional framework for developing the required critical technologies is as follows. First, meta-learning and meta-reasoning capabilities—learning how to learn and reason—are foundational. Rather than relying on static training, agents continuously refine their internal learning processes to autonomously adapt to new tasks and conditions. This allows the system to create specialized sub-agents for various operational tasks and teaches them optimal strategies. Second, replication logic creates next-generation agents to improve operational efficiency, reduce false-alarm rates, and field additional task-organized agents to address emergent or previously unanticipated requirements. Third, autonomous resource management mechanisms must dynamically replicate, allocate, sustain, scale, or retire agents to maintain operational efficiency and mission alignment. Finally, secure, low-probability-of-intercept/low-probability-of-detection (LPI/LPD) ad hoc networking and communications are essential to enable mission-focused teaming, decentralized coordination, and Reachback to share operational insights across agent networks beyond local clusters. Critical SI technology components and developments are as follows:

• Self-reflecting Agentic SI system to provide: 1) an introspection feedback loop to enable the Agentic SI system to analyze its own plan, spot errors or weak reasoning logic, and revise them before acting, which would make the system better at goal-seeking and improve its strategy; and 2) a retrospection feedback loop to enable the Agentic SI system to analyze its past plans, decisions, and resulting outcomes. Spot and learn from both the advantages and the shortcomings of those past decision processes to improve future planning and reasoning logic.
• Identification of novel and evolutionary behaviors appearing in a new generation of SI-agent models.
• Methodology to specify how self-replication and coevolutionary processes are initiated and developed, strategic life-cycle planning, resource acquisition for autonomous self-replication, and survival tactics that exploit operational ecosystem.
• Contextual modeling: relational, spatial, logic-based, uncertainty, and hybrid context modeling.
• Meta learning: decision tree and multilayer perception classifiers, collaborative filtering, pattern mining, deep learning.
• Spatiotemporal synchronization methods for multimodal data across decentralized architectures.
• Multimodal contextual signal processing and fusion.
• Cross-domain contextual collaborative learning, inference, and recognition.
• Meta reasoning, contextual reasoning, and risk assessment and reduction.
• Contextual priority-based task management balancing operational objectives such as persistence, endurance, and collections.

The scope of the initial feasibility study and deliverables are as follows: a) define the challenge problem that will likely meet the benchmark performance metrics; b) review a set of candidate Agentic SI methods and data requirements; c) develop a low-cost initial proof-of-concept prototype for the assessment; d) evaluate feasibility against measurable criteria; and e) deliver a proposed Phase II prototype development plan including technical risk reduction criteria and next steps.

Work produced in Phase II may become classified. Note: The prospective contractor(s) must be U.S. owned and operated with no foreign influence as defined by 32 U.S.C. § 2004.20 et seq., National Industrial Security Program Executive Agent and Operating Manual, unless acceptable mitigating procedures can and have been implemented and approved by the Defense Counterintelligence and Security Agency (DCSA) formerly Defense Security Service (DSS). The selected contractor and/or subcontractor must be able to acquire and maintain a secret level facility and Personnel Security Clearances. This will allow contractor personnel to perform on advanced phases of this project as set forth by DCSA and ONR in order to gain access to classified information pertaining to the national defense of the United States and its allies; this will be an inherent requirement. The selected company will be required to safeguard classified material during the advanced phases of this contract IAW the National Industrial Security Program Operating Manual (NISPOM), which can be found at Title 32, Part 2004.20 of the Code of Federal Regulations.

PHASE I: Determine the technical feasibility of designing and developing the Agentic SI Swarm technology as explained in the Description. Draw key distinctions for the proposed design approach compared to the current state-of-the-art naval ISRT information exploitation cycles. Ensure that the proposed design concept must clearly define the steps that naval ISRT platforms and joint services can capitalize on for transformational SI-assisted ISRT and engagement operations with dominant applications within the next five years. Motivate the design with a compelling challenge problem supported by relevant datasets. Consider a challenge problem that corresponds to cross-domain littoral operations and navigational risks associated with hostile anti-access/denied-access enforcement scenario that considers seaside terrain, weather, the maritime order of battle and maneuvers; engagement rules and doctrine; logistics and supply demands, etc. Hostile enforcement may include hypersonic missiles, ballistic missile attacks, sea-skimming cruise missile attacks, anti-submarine track and trail, torpedo attacks, littoral mining, and swarming weapons. Testing and demonstrations may use a mix of open-source intelligence (OSINT) datasets, synthetic datasets from DON, MCML, AIS maritime traffic, commercial satellite imagery, or similar sources. Conduct Agentic SI Swarming performance assessment, including cause and effect sensitivity analysis on contextual understanding; confidence rate on SI-generated engagement plans, options, and risk reduction associated with the ups and downs of encounters; confidence rate on SI-risk escalation controls enabling trust decisions across the integrated chain-of-command; efficiency gains in human responsiveness through timely decision-making, chain-of-actions, and resources spent.

Performance metrics for Agentic SI system in a controlled test environment are as follows: task success and completion rates: 90%+; goal attainment rate: 90%+; tool selection accuracy: 95%+; plan adherence: 90%+; hallucination rate: under 2%; safety violation rate: 0% in policy tests (any policy-breaking action is a hard fail); prompt-injection resistance: 95%+ success against red-team prompts (measure on adversarial test sets); human override rate: < 5% on routine tasks; recovery rate: 80%+; end-to-end latency: under 60 seconds for interactive flows; replication success rate: 100% only when explicitly authorized.

Policy, safety, and security performance metrics that are controlled by the embedded SI Guardrail-Enforcers are as follows:

• For self-replicating Agentic SI system to ensure the system can perform useful work without uncontrolled replication, drift, or excessive human intervention. In other words, it must replicate when authorized, and never on its own outside policy. The performance metrics include goal attainment rate: 90%+; tool selection accuracy (i.e., preventing cascading failures from bad tool use): 95%+; hallucination rate: under 2%; human override rate: under 5%; recovery rate: 80%+; boundary violation rate: 0%; and Uncontrolled replication rate: 0%.
• For Hybrid-evolving Agentic SI system, the performance metrics include improvement gains: greater than 20% (self-evolution should improve performance on unchanged tasks); regression rate: less than 5% on core tasks (new versions should not break stable capabilities); reproducibility (evolved behavior must remain traceable and can be validated): 95%+; and Drift from policy, safety, and security: 0% -- self-evolution cannot change the operational constraint envelope.

Assess the feasibility of applying the Agentic SI methods to a candidate challenge problem. Perform requirements analysis, data and interface assessment, evaluation of candidate SI approaches, and development of a limited proof of concept to determine technical viability. Produce a feasibility report, preliminary architecture, risk assessments, and a proposed Phase II prototype development plan.

PHASE II: Conduct proof-of-concept and prototype development incorporating the recommended candidate technology from Phase I. Demonstrate the operational effectiveness based on the following criteria: (a) prioritized sensor alerts, (b) prioritized threat escalation, (c) measured severity of events, and (d) measured analytic completeness. Test and demonstrate the improved capability based on the performance metrics detailed for Phase I. Provide analytics, SI tools, models, prototypes, T&E and demonstration results, and interface requirements. Develop a final report to include a detailed design of the system and a plan for transition to the program of record in Phase III.

Work in Phase II may become classified. Please see note in the Description.

PHASE III DUAL USE APPLICATIONS: Advance these capabilities to TRL-7 and integrate the technology into the Maritime Tactical Command and Control POR, Marine Air-Ground Task Force Command and Control, or ISR processing platforms at the Marine Corps Information Operations Center.

Once conceptually and technically validated, demonstrate dual-use applications of this technology in civilian law enforcement and security services.

REFERENCES:

1. Finn, C.; Abbeel, P. and Levine, S. "Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks." IEEE Internet of Things Journal Vol: 5 (Issue: 1), November 2017, pp. 1-27. https://arxiv.org/abs/1703.03400
2. Huisman, M.; van Rijn, J.N. and Plaat, A. "A survey of deep meta learning." Artificial Intelligence Review, April 2021. https://arvix.org/2010.03522
3. Wang, S.; Chen, K.; Yu, M.; Zhao, P. and Duan, H. "Multimodal meta-adaptive reasoning network with dynamic causal modeling and co-evolution of quantum states." Scientific Reports, July 2025, pp 1-19. https://www.nature.com/articles/s41598-025-1127-x
4. Breeden, Joseph L. "The evolution of goals in AI agents." Springer, Published online: March 2025. https://link.springer.com/article/10.1007/s43681-025-00691-y
5. Wang, X.; Feng, T.; Liu, H. and Zhu, W. "Autonomous Embodied AI, Towards Self-evolving Intelligence", Springer, 2026. https://springer.com/book/10.1007/978-981-95-7750-7
6. "National Industrial Security Program Executive Agent and Operating Manual (NISP), 32 U.S.C. § 2004.20 et seq. 1993". https://www.ecfr.gov/current/title-32/subtitle-B/chapter-XX/part-2004

KEYWORDS: Self-Replicating; Hybrid-Evolutionary; Self-Evolving; Coevolving; Agentic; SI; Swarm; Autonomous; Guardrail; Enforcers

TPOC 1: Allen Moshfegh
allen.moshfegh.civ@us.navy.mil

TPOC 2: Lodewijk Brand
lodewijk.w.brand.civ@us.navy.mil