The Doctrine Gap: DARPA's DICE Programme and the $54 Billion Question.

DARPA's Information Innovation Office published Special Notice DARPA-SN-26-72 on 28 April 2026, formally inviting industry concepts for the Decentralized Artificial Intelligence through Controlled Emergence programme. The Special Notice response deadline closes today, 19 May 2026, with a Proposers Day scheduled for 29 May 2026 at the Executive Conference Center in Arlington, Virginia. Programme manager Susmit Jha, in his published programme synopsis, frames the technical objective in three sentences: develop the theory and algorithms for decentralised coordination and local inference control; enable a scalable, adaptive, resilient collective of heterogeneous AI agents; sustain long-time-horizon missions in contested environments while remaining under human control.

The conceptual centre of the programme is what DARPA terms controlled emergence: complex system-level behaviour arising from simple local rules, but constrained and predictable. Each agent operates without a central control node yet coordinates through a peer-to-peer network. The architecture must remain on mission, maintain doctrine, suppress misbehaviours, and remain resilient to agent loss or compromise. Critically, DICE is also designed to be resilient to rogue AI agents that might develop misaligned instrumental goals. That last specification is the most strategically significant. It is the first public DARPA requirement we are aware of that explicitly addresses the problem of internal AI-agent misalignment as a planning constraint, rather than treating it as a post-deployment safety concern.

The programme scope is deliberately bounded. DARPA states explicitly that DICE will not include the development and deployment of autonomous systems in the real world. The work will demonstrate DICE architectures in simulation environments, targeting measurable gains in scalability, adaptability, and resilience against both benign failures and adversarial attacks. The simulation-only constraint is consistent with DARPA's pattern on architecture programmes: define the requirement and the architecture at the agency, hand off the productionisation to follow-on procurement vehicles managed elsewhere in the Pentagon's autonomous warfare apparatus.

DARPA's Microsystems Technology Office released Request for Information DARPA-SN-26-76 on Materials for Physical Compute in Untethered Robotics in April 2026, with a response deadline of 27 May 2026 and a follow-up workshop scheduled for June or July 2026. The RFI is the architectural companion to DICE. Where DICE addresses how a swarm coordinates, the Materials RFI addresses how each individual agent thinks. The technical premise is that conventional CPU and GPU-centric architectures impose a fundamental energy and latency penalty on small autonomous platforms, particularly in communication-denied environments. By embedding sensing, actuation, and local inference directly into the same physical substrate, the latency between sensing and response can be reduced to material-level timescales, and the energy budget can be reallocated from data movement to flight or operational duration.

DARPA's stated areas of interest are specific and exclusionary. The MTO is interested in materials that integrate sensing, actuation, and elements of control into a unified substrate; in dynamic adaptive closed-loop compute that can perform inference without centralised processors; and in technologies that enable real-time decision-making with minimal latency and reduced power demands. DARPA explicitly states that incremental hardware improvements, software-only AI approaches, and conventional CPU or GPU-centric architectures are not of interest under this RFI. The exclusion is as informative as the inclusion: DARPA is publicly committing to a thesis that the next autonomy substrate will not be downsized versions of today's compute stack.

Read alongside DICE, the Materials RFI completes a coherent technical agenda. DICE provides the collective intelligence layer; the Materials RFI provides the individual-agent substrate. Together they target the two technical preconditions for the containerised 500-aircraft constellation specified in DARPA-SN-26-33 a month earlier: the autonomy software that coordinates the swarm, and the hardware substrate that lets each platform operate in a GPS-denied electromagnetically contested environment without continuous human supervision.

The clearest critique of the FY27 autonomous warfare budget came in mid-May 2026 from retired Gen. David Petraeus, former Director of the CIA, and scholar Isaac Flanagan in an opinion piece in The Hill. Petraeus and Flanagan argue that the technical and procurement architecture is being built faster than the institutional structure required to use it effectively. Their central numeric claim, attributed to their own analysis, is that less than 2 percent of the Pentagon's $54 billion FY27 autonomous warfare request is being spent on doctrine, training, and force design. Their recommendation: Congress should allocate at least 5 percent of the autonomous warfare budget to these institutional foundations before further procurement expansion.

The historical anchor of their argument is the Predator combat air patrol. According to their analysis, each continuous Predator surveillance orbit required approximately 150 personnel: pilots, sensor operators, communications technicians, armaments specialists, imagery analysts, maintenance crews, operations planners, linguists, and additional intelligence professionals. The limiting factor on US drone coverage in the early Iraq and Afghanistan campaigns was not the number of aircraft. It was the trained personnel and the organisational structure to operate them. The autonomous-warfare equivalent, they warn, is that the Pentagon risks repeating the early-2000s mistake by buying platforms and software faster than it builds the doctrine and trained workforce to integrate them. Their formulation, also published in The Hill, is that without joint doctrine, 'autonomous' is just a label, not a capability.

DICE is the technical answer to the operator bottleneck, but only if institutional doctrine catches up. A 500-aircraft autonomous constellation operating at constellation-scale Autonomy Level 4 requires rules of engagement that have not been written, targeting authorities that have not been delegated, and cross-service protocols for human-in-the-loop versus human-on-the-loop control that the Pentagon has not yet settled. Without that doctrine, the technical capability cannot be deployed at the scale the FY27 budget assumes. The institutional gap is the dominant execution risk of the autonomous warfare buildout, and it is a risk that does not appear on any vendor's product roadmap because no vendor can fix it.