DARPA DICE Program Targets Decentralized Control of AI Agent Swarms

DARPA's DICE program, BAA HR001126S0010, seeks decentralized coordination algorithms for heterogeneous AI agent swarms, with a budget start month of 1 February 2027. The FY2027 request for the Defense Autonomous Warfare Group is $54.6 billion, about 240 times the FY2026 figure.

Program snapshot

DICE's stated goal is to develop theory and algorithms for decentralized coordination and local reasoning control so scalable, adaptive, high-resilience swarms of heterogeneous AI agents can autonomously execute long-duration tasks in adversarial environments while remaining under human control. Program manager Dr. Susmit Jha joined DARPA in August 2025. He was previously technical director of the Computer Science Laboratory at SRI International and head of its neuro-symbolic computing and intelligence research group. He was principal investigator on DARPA Trusted Neuro-symbolic Learning and Reasoning (ANSR) and other efforts, with more than $26 million in funding, and led related work in DARPA's Information Science and Technology (ISAT) advisory group.

A program preview (DARPA-SN-26-65) was issued on 21 April 2026, a proposer-day notice (SN-26-72) on 28 April, and a hybrid proposer day on 29 May in Arlington, Virginia. Broad Agency Announcement HR001126S0010 was released on 10 June. Abstracts were due 30 June and full proposals 25 August. The budget baseline start month is 1 February 2027. Multiple awards are expected as procurement contracts, prototype other-transaction (OT) agreements, research OT agreements, and cooperative agreements. Offerors must hold CMMC Level 1 certification at submission. As of 19 July 2026 the program was still in proposal preparation and had no selected performers.

*Table 1. DICE program basic elements*

Budget and organization

The 2023 Replicator initiative aimed to field thousands of attritable autonomous systems in the Indo-Pacific. It was dissolved at the end of 2025 and folded into the Defense Autonomous Warfare Group (DAWG) after platform immaturity, command-and-control integration problems, lack of a dedicated budget line, and failure to buy software able to orchestrate large heterogeneous drone swarms. The Trump administration FY2027 budget request listed $54.6 billion for DAWG, compared with $225.9 million in FY2026, about a 240-fold increase. About $1 billion sits in the regular base budget and about $53.6 billion in a flexible reconciliation pool over five years. That item is about 15 percent of a $350 billion defense reconciliation package and exceeds the U.S. Marine Corps annual budget. Defense Secretary Hegseth announced an autonomous operations deputy unified command. U.S. Southern Command stood up a Southern Autonomous Warfare Command (SAWC) to work with DAWG.

DAWG's Drone Dominance effort, about $1.1 billion, is described as buying about 300,000 domestically produced armed drones by the end of 2027. Anduril received a 10-year Lattice command-and-control contract of about $20 billion in late 2025. DICE is presented as addressing how thousands of reasoning AI agents can act without a central node and remain inside human control. Since the Predator era, each MQ-9 has required more than 180 personnel.

Technical premise and lineage

The BAA states that future conflict will run at machine speed and that centralized planning and scripted orchestration are too slow, too rigid, and too easy to predict. Mainstream multi-agent systems still use a central orchestrator to split tasks, assign roles, and merge results. Cited problems include a single point of failure under electronic warfare, network attack, and communications loss; a cognitive bottleneck as agent count and interaction complexity grow past base-model context and reasoning limits; and predictable pre-orchestrated behavior. The BAA cites the internet, Conway's Game of Life, and animal groups as local-rule references, while noting DICE agents may reason, plan, call tools, remember, and generate intermediate goals.

The 2014 Collaborative Operations in Denied Environment (CODE) program showed one human mission commander supervising multiple drones through find-to-engage, with Raytheon completing Phase 3 software. The 2017 Offensive Swarm-Enabled Tactics (OFFSET) program scaled to 250 air and ground unmanned platforms, with Raytheon BBN and Northrop Grumman as integrators. In late 2021 at Fort Campbell, a single operator controlled 130 physical plus 30 virtual drones (Raytheon) or 174 platforms (Northrop Grumman) for 3.5 hours. Air Combat Evolution (ACE) recorded a 5-0 AI win over an F-16 pilot in the 2020 AlphaDogfight Trials. That work fed the Collaborative Combat Aircraft (CCA) program. Anduril's YFQ-44A and General Atomics' YFQ-42A won CCA, with a first production lot of 150 aircraft in June 2026.

*Table 2. DARPA technical lineage for DICE*

A January 2026 Hegseth coordination memo set an AI-first strategy and seven priority sprint projects. Swarm Forge, Agent Network, and Ender's Foundry align with DICE. The memo requires the Chief Digital and Artificial Intelligence Office to deliver the latest frontier models to operators within 30 days of public release. DISCORD, under Ender's Foundry, posted a solicitation in May 2026 to generate diverse tactical option sets for commanders.

Competition, doctrine, and vendors

U.S. public assessments cite Chinese production of more than 100,000 military-grade drones a year. In March 2026, CCTV Military showed the domestic Atlas drone swarm combat system. The Jiutian large UAV swarm mothership is described as completing target ranking and path planning in 0.3 seconds. Vendors including Aerospace CH have claimed 100-aircraft-class control by one operator, centerless networking, and tactical recombination under 0.3 seconds. A 2026 National Defense University unmanned-operations outlook said command is moving from hierarchy toward decentralization, with commanders setting mission bounds rather than detailed orders.

Retired Gen. David Petraeus and Flanagan wrote in The Hill that of more than $54 billion in FY2027 autonomous-operations funding, the share for doctrine, training, and force design is under 2 percent, and they argued it should be at least 5 percent. Department of Defense Directive 3000.09 requires an appropriate level of human judgment for autonomous weapon systems.

In February 2026 Anthropic was labeled a supply-chain risk by Hegseth after holding lines against large-scale domestic surveillance and fully autonomous lethal weapons. Anthropic sued in federal court and won a preliminary injunction on 26 March. OpenAI later took a Pentagon classified deployment agreement of about $200 million. By 1 May the Department of Defense had signed classified-network AI agreements with SpaceX (xAI), OpenAI, Google, NVIDIA, Reflection, Microsoft, and Amazon AWS. Anthropic was excluded.

Three-phase plan

The BAA describes controlled emergence: intelligence may emerge from heterogeneous agent interaction, but architectural controls are meant to keep doctrine, suppress misbehavior, hold task focus, and support long-horizon coordination. Control is framed as local inference-time bounds rather than a top-down command stream. Global norms may be preset or arise spontaneously if they act only through local adapters without a fixed central authority. Leader election and dynamic hierarchy are in scope if they adapt quickly and resist disruption.

The 36-month effort has three stages. Phase 1 lasts 9 months at about 500 agents and 5,000 interactions, testing decentralized methods against a centralized multi-agent baseline. Phase 2 lasts 15 months at 5,000 agents and 50,000 interactions, adding agent failure, deceptive inputs, and subverted or uncontrolled members. Inter-team playoffs are set for months 13, 16, and 19, with a month-23 playoff to select Phase 3 teams. Phase 3 lasts 12 months at 100,000 agents and 1 million interactions, adds agents that create other agents, and ends with a final contest at month 35.

TA3 performers must release the Phase 1 test and evaluation platform final version by month 5 and issue a centralized baseline to all TA1 and TA2 teams by month 6, with updates allowed through final evaluation. Simulation frames roll from adaptation to resistance to scale. Program meetings are planned in Washington, San Francisco, and Boston.