DARPA DICE (HR001126S0010): What Startups Need to Know
Below is a brief summary. Please check the full solicitation before applying (link in resources section).
Quick Answer
DARPA's Information Processing Techniques Office (IPTO) has released the Decentralized Artificial Intelligence through Controlled Emergence (DICE) Broad Agency Announcement, solicitation number HR001126S0010. DICE funds research into decentralized multi-agent AI systems that can coordinate and stay under control without a central orchestrator, for use in long, complex, contested missions. The program runs 36 months across three phases. Abstracts are due June 30, 2026, and full proposals are due August 25, 2026. Multiple awards are anticipated, using Procurement Contracts, Other Transaction Agreements, and Cooperative Agreements. There is no set-aside for small businesses, but all responsible U.S. and non-U.S. sources may compete.
Program Overview
DICE aims to move multi-agent AI systems away from today's brittle, centrally orchestrated model (where a single planner, router, or "boss agent" assigns tasks and aggregates results) and toward a decentralized architecture where heterogeneous AI agents self-organize, coordinate peer-to-peer, and stay aligned to mission intent without a central point of control or failure.
The core problem DARPA is trying to solve: current multi-agent systems (built on frameworks like LangGraph, CrewAI, or AutoGen) don't scale well, adapt slowly to changing conditions, struggle with incomplete or conflicting information, and are fragile to adversarial manipulation, since a single compromised orchestrator can take down the whole system. DICE wants a fundamentally different architecture, not an incremental improvement on existing orchestration tools.
The technical solution DARPA envisions is a "local adaptor" that plugs into any AI agent (regardless of vendor or model) and handles two jobs at once: peer-to-peer coordination with other agents, and local control that keeps that individual agent aligned to its role and mission over long stretches of time, even under adversarial pressure.
DARPA explicitly wants revolutionary advances, not evolutionary improvements to existing practice. Proposals that read as generic, boilerplate, or broadly aggregated will be found nonconforming before they're even substantively reviewed.
Who This Is For
This solicitation is a strong fit for organizations working in:
Multi-agent AI systems and agent orchestration
Distributed systems, consensus algorithms, and fault-tolerant computing
Reinforcement learning, particularly multi-agent RL
Mechanistic interpretability, activation steering, and representation engineering
Game theory and mechanism design
Formal methods and control theory
AI red-teaming, adversarial robustness, and Byzantine fault tolerance
Simulation environment and test infrastructure development for agentic systems
It is not a fit for teams proposing incremental improvements to existing orchestration frameworks, narrow cybersecurity network-simulation tools, or small-scale robotics work built around high-fidelity physics engines. DARPA calls both of those out explicitly as out of scope for the Test and Evaluation area.
The Three Technical Areas
DICE is structured around three Technical Areas (TAs). Proposers pick one lane: TA1 and TA2 must be proposed together as a combined effort, while TA3 is proposed on its own. Organizations selected to perform TA3 are not eligible to also perform TA1 or TA2.
TA1: Self-Organization via Peer-to-Peer Coordination and Distributed Consensus TA1 teams build the theory and algorithms that let AI agents form teams, divide up tasks, and reach consensus without a central planner. Approaches can include distributed auctions, multi-agent reinforcement learning, game-theoretic mechanism design, and Byzantine fault-tolerant consensus. Success is measured by scalability (how many agents, roles, and messages the system can handle), adaptability (how fast it recovers from failure or disruption), and resilience (how well it holds up under conflicting, deceptive, or adversarial information).
TA2: Role Coherence and Local Inference Control TA2 teams build inference-time control methods that keep individual AI agents faithful to their assigned role and mission over long conversations and many reasoning steps, without turning them into rigid, low-creativity systems. Approaches can include activation steering, representation engineering, memory editing, context pruning, and game-theoretic incentive design for closed, black-box models where internal activations aren't accessible. TA2 must work for both open-weight models and proprietary black-box models.
TA3: Test and Evaluation TA3 performers build the simulation environment, the defense-relevant use cases, and the evaluation infrastructure that everyone else's technology gets tested against. This includes building a baseline centrally orchestrated multi-agent system for Phase 1 comparison, running team-vs-team playoffs in Phases 2 and 3, and standing up a platform that can eventually simulate 100,000 agents exchanging a million messages. TA3 work can involve Controlled Unclassified Information (CUI) and requires demonstrated CUI safeguarding capability as a condition of passing the initial conformance review.
Program Structure and Scale
DICE runs 36 months across three phases, with mission complexity scaling up dramatically at each stage:
Phase 1, Decentralization (9 months): Establish that a decentralized approach beats a centrally orchestrated baseline. Target scale is 500 agents and 5,000 interactions between them. Evaluated head-to-head against a DARPA-built SOTA orchestrated baseline in month 8.
Phase 2, Adversarial Robustness (15 months): Shift focus to surviving Byzantine failures, deception, and compromised agents. Target scale is 5,000 agents and 50,000 interactions. Evaluated through team-vs-team playoffs in months 13, 16, and 19, with a final playoff in month 23 that down-selects which TA1/TA2 teams advance to Phase 3.
Phase 3, Scale (12 months): Push the architecture to its limits. Target scale is 100,000 agents and 1 million interactions, with agents beginning to create other agents. Ends with a final team-vs-team competition in month 35.
For budgeting purposes, proposers should treat February 1, 2027, as the Month 1 start date, even though actual meeting locations and scheduling are notional and subject to change.
Funding Structure
DARPA has not published a specific dollar ceiling or per-award funding range in this BAA. What is specified:
Multiple awards are anticipated
Award instruments may include Procurement Contracts, Other Transaction (OT) Agreements for Prototype (10 U.S.C. Section 4022), OT Agreements for Research (10 U.S.C. Section 4021), and Cooperative Agreements
The Government Contracting Officer has sole discretion over which instrument type is used, regardless of what a proposer requests
Resource sharing may be required for OT Prototype or OT Research agreements under 10 U.S.C. Sections 4021 and 4022
Cost realism is an official evaluation criterion, meaning proposed budgets need to be defensible and tied directly to the proposed labor hours, materials, and technical approach, not just a round number
Because DARPA intends to retain Government Purpose Rights in all deliverables and expects most DICE research to be open-sourced, proposers who do not intend to open-source their work need to explain why in the proposal itself.
Key Dates
Proposers Day: May 29, 2026
Posting Date: June 10, 2026
Abstract Due Date: June 30, 2026, 2:00 PM ET (strongly encouraged, not mandatory)
Question Submittal Closes: August 18, 2026, 2:00 PM ET
Full Proposal Due Date: August 25, 2026, 2:00 PM ET
Submitting an abstract first is optional but strongly recommended. DARPA reviews abstracts in two steps: a strict technical conformance check, then a comprehensive review. Every proposer who submits an abstract gets a letter back marked Nonconforming, Encouraged, or Discouraged, each with DARPA's rationale attached. A Discouraged rating doesn't legally bar you from submitting a full proposal, but it's a strong signal about your odds.
Eligibility
Open to all responsible U.S. and non-U.S. sources, including small businesses, Historically Black Colleges and Universities, and minority institutions, though there is no funding set-aside for any of these categories. Non-U.S. participants must comply with applicable nondisclosure agreements, security regulations, and export control laws.
Not eligible: Federally Funded Research and Development Centers (FFRDCs), University Affiliated Research Centers (UARCs), and government entities including national labs. UARCs and FFRDCs interested in the program should contact the DARPA point of contact before the abstract deadline to discuss joining as part of the government team instead.
Organizations cannot simultaneously act as a technical performer and provide Systems Engineering Technical Assistance, Advisory and Assistance Services, or similar support services on this program, unless the DARPA Deputy Director grants a written waiver.
Security and Compliance Requirements
TA1 and TA2 work is expected to be entirely unclassified. TA3 work spans both unclassified and Controlled Unclassified Information (CUI), and TA3 proposers must demonstrate CUI safeguarding capacity as part of the mandatory conformance review, or they will be found unacceptable before technical merit is even evaluated.
CMMC (Cybersecurity Maturity Model Certification) applies to any resulting procurement contract. CMMC Level 1 certification is required at the time of proposal submission. Level 2 applies to any work touching CUI. A current CMMC self-assessment must already be recorded in the Supplier Performance Risk System (SPRS) before award, and a valid, current CMMC certification is a hard condition of award, meaning proposers without it are ineligible to be awarded regardless of technical score. Contracts awarded after November 10, 2026 that involve CUI will additionally require a CMMC Third-Party Assessment on file in SPRS before award.
Because some fundamental research under this program may disclose performance characteristics of military systems considered unique and critical to defense, some awards will carry publication restrictions requiring DARPA's permission before releasing results, even though DARPA's default policy favors unrestricted publication of fundamental research.
How Proposals Get Evaluated
Evaluation happens in two steps.
Step 1, Mandatory Technical Conformance: DARPA checks whether the technical approach is genuinely original. Proposals that read as generic, boilerplate, or broadly aggregated without mission-specific synthesis are found nonconforming and go no further, regardless of everything else in the proposal. TA3 proposals face an additional conformance check on demonstrated CUI safeguarding capacity.
Step 2, Comprehensive Evaluation: Conforming proposals are then scored against five criteria, listed here in descending order of importance:
Overall Scientific and Technical Merit: Is the approach innovative, feasible, and complete, with clear tasks, deliverables, identified risks, and a realistic schedule?
Technical Qualifications: Does the team have the expertise and demonstrated commitment to actually execute?
Potential Contribution and Relevance to the DARPA Mission: Does this create or prevent technological surprise and strengthen the national security technology base?
Plans and Capability to Accomplish Technology Transition: Is there a credible path to transition the technology, and does the proposed IP structure help or hurt the government's ability to transition it?
Cost Realism: Are the proposed costs defensible, tied to actual labor hours and materials, and does the approach make efficient use of prior related research already available?
What You Need to Submit
Abstract (optional but strongly encouraged): Submitted through the DARPA BAA Tool (BAAT) using the required A1 Abstract Template. Anything not explicitly requested in the template will not be reviewed, so following the format matters as much as the content.
Full Proposal: Also submitted through BAAT, and must include:
P1, Volume I Proposal Template (Technical and Management)
P2, Volume II Proposal Template (Cost)
P3, Proposal Summary Slide Template
P4, DARPA Standard Cost Proposal Spreadsheet
Associate Contractor Agreement (for informational purposes)
If requesting an OT Agreement or Cooperative Agreement, key personnel must also submit a CF Biographical Sketch and CF Other Support form, using NSF's instructions for completing the Common Disclosure Forms.
DARPA strongly encourages proposers to submit red-lined, track-changes edits to the Model Contract or Model OT Agreement showing exactly what terms they would want to negotiate if selected, with comment boxes explaining the reasoning behind each requested change.
Program Meetings
Month 1: Program Kickoff Workshop, Washington D.C.
Month 5: Phase 1 PI Meeting, San Francisco, CA
Month 10: Phase 2 Kickoff Workshop, Boston, MA
Month 18: Phase 2 PI Meeting, Washington D.C.
Month 24: Phase 3 Kickoff Workshop, Washington D.C.
Month 30: Phase 3 PI Meeting, San Francisco, CA
Month 35: Final PI Meeting and Demo, Washington D.C.
Locations are notional and subject to change.
Why This Matters Beyond the Award
DICE sits at the intersection of a real commercial trend and a real defense need. Every major AI lab is racing toward multi-agent systems right now, and almost all of them are built on centralized orchestration frameworks that DARPA is explicitly calling brittle and unfit for contested conditions. A company that builds a genuinely decentralized coordination or control architecture under DICE isn't just fulfilling a defense contract, it's building IP that's directly relevant to where commercial agentic AI is headed as multi-agent systems scale past what a single orchestrator can reasonably manage.
Frequently Asked Questions
What is DARPA DICE? DICE, or Decentralized Artificial Intelligence through Controlled Emergence, is a DARPA IPTO program funding research into multi-agent AI systems that coordinate peer-to-peer and stay aligned to mission intent without relying on a central orchestrator, so they can operate in contested, degraded-communication environments.
What is the solicitation number for DICE? HR001126S0010.
When are DICE abstracts due? June 30, 2026, at 2:00 PM Eastern Time. Abstracts are strongly encouraged but not required.
When are full DICE proposals due? August 25, 2026, at 2:00 PM Eastern Time.
How long is the DICE program? 36 months, split into three phases: 9 months for Phase 1 (Decentralization), 15 months for Phase 2 (Adversarial Robustness), and 12 months for Phase 3 (Scale).
How much funding does DICE provide? DARPA has not published a specific dollar ceiling in the BAA. Multiple awards are anticipated, and funding instruments may include Procurement Contracts, Other Transaction Agreements, and Cooperative Agreements. The government has sole discretion over which instrument type is used and negotiates terms individually with each selectee.
What are the three technical areas in DICE? TA1 covers peer-to-peer coordination and distributed consensus among agents. TA2 covers local inference-time control that keeps individual agents aligned to their role and mission. TA3 covers test and evaluation, including building the simulation environment and running competitive evaluations. TA1 and TA2 must be proposed together; TA3 is proposed separately, and TA3 performers cannot also perform TA1 or TA2.
Is DICE open to small businesses and non-U.S. companies? Yes. All responsible U.S. and non-U.S. sources may propose, including small businesses, Historically Black Colleges and Universities, and minority institutions, though there is no funding set-aside reserved for any of these groups. Non-U.S. participants must comply with applicable export control and security requirements.
Who is not eligible to propose on DICE? Federally Funded Research and Development Centers (FFRDCs), University Affiliated Research Centers (UARCs), and government entities including national labs cannot propose. These organizations can contact DARPA before the abstract deadline about joining the government team instead.
Does DICE require security clearances or CMMC certification? TA1 and TA2 work is expected to remain fully unclassified. TA3 work spans unclassified and CUI. Any resulting procurement contract requires CMMC Level 1 certification at the time of proposal submission, with Level 2 required for work involving CUI, and a valid CMMC certification is a hard condition of contract award.
What happens if my abstract is found nonconforming? DARPA will notify you that it was not reviewed in the comprehensive step and will explain its rationale. You can still submit a full proposal even after a Discouraged or Nonconforming abstract determination, but DARPA's feedback is a strong signal of how competitive your approach is likely to be.
Does DARPA retain rights to the technology developed under DICE? DARPA intends to retain Government Purpose Rights in all DICE deliverables across all three technical areas, including source code. DARPA also expects most DICE research to be open-sourced. If a proposer does not intend to open-source their work, they need to explain that rationale directly in the proposal.
What kind of use cases will DICE be tested against? DARPA-relevant use cases such as operational planning for conflict scenarios, resource acquisition, logistics, and battle management, tested in simulation environments rather than physical hardware. Explicitly out of scope are use cases centered purely on cybersecurity network simulation and small-scale robotics with high-fidelity physics engines.
