DARPA's DICE Program Bets $Millions That the Future of Military AI Is a Swarm That Governs Itself — Proposals Close August 25

Most of the AI money moving through the federal government in 2026 is chasing bigger models, faster inference, and better benchmarks. DARPA's newest AI program is chasing something almost orthogonal to all of it: how do you get a crowd of imperfect, semi-autonomous AI agents to behave like a disciplined unit — with no commander in the middle, in an environment actively trying to break them, and without a human babysitting every decision?

That is the question behind DICE — Decentralized Artificial Intelligence through Controlled Emergence — issued as Broad Agency Announcement HR001126S0010 by DARPA's Information Processing Techniques Office on June 10, 2026. Abstracts were due June 30. The full-proposal deadline is August 25, 2026 at 2:00 PM ET, with a final questions window closing August 18. If you are reading this at the start of the month, you have roughly three and a half weeks — and if you have not already submitted an abstract, you are behind. But DICE is worth understanding even if you cannot make this cycle, because it signals where a large slice of defense AI funding is heading: away from the single monolithic model and toward governed collectives.

What "controlled emergence" actually means

The phrase in the program's name is doing a lot of work. Emergence is what you get when many simple agents interact and produce collective behavior no single agent was programmed to perform — flocking, swarming, market pricing. It is powerful and it is cheap to scale. It is also, historically, uncontrollable: emergent systems drift, cascade, and do surprising things at the worst possible moment. The military cannot field a swarm that might spontaneously decide to do something off-doctrine.

The opposite pole is centralized control — one orchestrator that tells every agent what to do. That is predictable and auditable, but it has a single point of failure, it does not scale, and in a contested environment the adversary's first move is to sever or spoof the central node.

DICE is a bet that there is a third path: emergence you can steer. DARPA wants "a scalable, adaptive, and resilient collective of heterogeneous AI agents" that can "self-organize, allocate tasks, share information, and adapt" through peer-to-peer mechanisms — while remaining, in the program's language, on mission, on doctrine, and under human control. The agents govern themselves, but the space of behaviors they can emerge into is bounded by design. That is the whole trick, and it is genuinely hard.

The three technical areas — and the one requirement that trips people up

The BAA splits the work into three technical areas:

• TA1 — Decentralized Coordination and Consensus. The distributed-systems half: peer-to-peer task allocation, information sharing, dynamic team formation, and consensus among heterogeneous agents with no central authority. This is descended from decades of work on distributed consensus algorithms and self-organizing systems, now applied to agents whose "nodes" are themselves language and reasoning models.

• TA2 — Local Inference Control. The frontier-model half: keeping each individual agent role-coherent and mission-aligned across long-duration missions using techniques like activation steering, memory editing, and context engineering. This is the machinery that stops an agent from drifting off-role, being talked out of its objective, or being corrupted into misbehavior.

• TA3 — Testing and Evaluation. Simulation environments and assessment frameworks to measure whether a decentralized collective actually outperforms a traditional centrally-orchestrated one under adversarial pressure.

Here is the structural detail that decides who is even eligible to compete for the core of the program: a core-technology proposal must address TA1 and TA2 together. You cannot bid TA1 alone as a distributed-systems shop, and you cannot bid TA2 alone as an AI-alignment lab. TA3 competes separately. That single sentence is the most important line in the solicitation, because it defines the team you have to build.

Why this is a team-formation problem, not a proposal-writing problem

The reason the TA1+TA2 requirement matters so much is that the two halves of DICE live in different research communities that rarely publish together. Distributed consensus and multi-agent coordination is a robotics, controls, and systems-theory tradition. Activation steering and inference-time control of foundation models is a 2024–2026 interpretability-and-alignment tradition. Very few organizations have deep benches in both.

That mismatch is the opportunity. The teams most likely to win are not the ones with the single strongest component — they are the ones that credibly integrate a distributed-systems group with a frontier-model-control group and can show the seam between them is engineered, not stapled. If you are a university lab strong in multi-agent systems, your fastest path to a competitive proposal is a co-PI who does mechanistic control of large models — and vice versa. If you are a startup, the same logic applies to a subcontract or teaming agreement. DARPA program managers read past the individual technical volumes to ask a blunt question: does this team have both halves, and have they thought about how the halves talk to each other?

The program shape tells you how to scope the work

DICE is 36 months, three phases, simulation-only. That structure carries planning signals worth reading carefully:

• Phase 1 asks you to demonstrate advantages over current state-of-the-art centrally-orchestrated systems. Your near-term milestones need to be about measurable superiority, not just a working prototype.
• Phase 2 turns adversarial: robustness against compromised agents and deliberately deceptive information injected into the collective. Build your architecture from day one assuming some fraction of your own agents are lying or captured.
• Phase 3 scales the collective to larger agent populations, testing whether coordination holds as numbers grow.

"Simulation-only" is a gift for non-traditional performers. There is no hardware integration, no flight test, no field deployment to fund or clear — which lowers the capital barrier and makes this genuinely accessible to small research teams and startups, not just primes with test ranges. The evaluation lives entirely in TA3's simulation environments, so the currency of the program is algorithms, theory, and reproducible experiments. That is a profile that favors sharp, focused teams.

Who should be looking hard at DICE

Eligibility is broad in the DARPA fashion — "all responsible sources capable of satisfying the Government's needs may submit a proposal." In practice, the strongest fits are:

• University labs in multi-agent systems, distributed AI, or control theory that can pair with an interpretability/alignment group.
• AI startups working on agentic frameworks, agent orchestration, or inference-time model control — DICE is a rare federal program that pays for exactly the multi-agent problems the commercial agent ecosystem is also racing to solve, with the added twist of adversarial robustness.
• Nonprofit research institutes with both systems and ML depth.

If you build agentic AI commercially, note the strategic overlay: the coordination, role-coherence, and anti-drift problems DICE funds are the same ones that make production multi-agent systems reliable. A DICE award is dual-use research capital for a capability the private market wants anyway.

What to do in the next three weeks — and what to do if you miss it

If you submitted an abstract and got encouragement, the full proposal is a sprint: lock your TA1+TA2 teaming now, use the August 18 questions deadline to resolve any ambiguity with the program office in writing, and build your Phase 1 milestones around demonstrable superiority over centralized baselines. Direct questions to DICE@darpa.mil.

If you cannot make August 25, treat DICE as a map. DARPA's I2O has planted a flag on governed decentralized agent collectives, and programs like this rarely arrive alone — follow-ons, related BAAs, and SBIR topics tend to cluster around a program office's active themes. Start building the TA1+TA2 partnership now so you are not assembling a team from scratch when the next solicitation lands. In defense AI, the teams that win are almost never the ones that started when the BAA dropped.

DICE is a wager that the next leap in military AI is not a smarter agent but a better-governed crowd of them. Whether or not that wager pays off, it is telling you where the money is looking. For the federal AI funding landscape this shift sits in, see our coverage of how AI now drives federal grantmaking, and browse active defense-innovation solicitations in the Granted grant database.