Many-on-many intercept is math before it is hardware
Strip away the airframes and a counter-swarm engagement is an assignment problem. N ready interceptors, M inbound tracks. Each pairing has a feasibility answer (can this interceptor reach that track with its battery, speed, and altitude envelope, with reserve to spare) and a cost (energy spent, magazine depth consumed, coverage surrendered). Add layered fallbacks: cheap interceptors first, mid-layer missiles for leakers, high-end rounds held for what threatens the asset you are defending. The solver has seconds, and the answer changes every time a track maneuvers or an interceptor falls out.
The Marine Corps Commandant stated the requirement directly: 'We've got to automate our drones so we launch them in swarms and they're talking to each other.' The CDAO's Swarm Forge solicitation and its Crucible demonstration ask for the same thing: end-to-end autonomous completion of mission sets. The services know the requirement. The open question is the architecture that computes the assignment.
The centralized solver dies with the link
The textbook answer is a central node that collects every track, computes a globally optimal assignment, and pushes tasking to each interceptor. It produces the best answer on a whiteboard and fails on contact. The central node is a single point of failure. The links carrying tracks up and tasking down are exactly what adversary EW attacks first. And a swarm launched to saturate your defense is also saturating your spectrum.
The decentralized alternative is an auction. Each interceptor scores the tracks it can feasibly reach using local information, bids, and resolves conflicts peer to peer over the mesh. No node is essential. Lose an interceptor and its bids disappear; the survivors re-bid and the coverage hole closes. Degrade the mesh and the auction runs on stale data with wider error bars instead of not running at all. The math is decades old. What changed is that the compute to run it now fits on the node itself.
We built a simulator to pressure-test the logic
EdgeLance built a counter-swarm assignment simulator to test decentralized coordination before any hardware flies. It runs three planner modes side by side: distributed auction, nearest-feasible greedy, and receding horizon. Four comms conditions: permissive mesh, degraded mesh, denied with stale tracks, and passive inference, where nodes bid from observed teammate behavior instead of received messages. Interceptor profiles from light sprint to heavy endurance, battery reserve floors, threat altitude bands, and five engagement policies from conserve-interceptors to protect-the-high-value-asset.
Every run produces a full accounting: which interceptor took which track and why, cost per engagement across the low layer, mid-layer missiles, and high-end reserve, and an event log of every bid, conflict, and commit. A Monte Carlo mode sweeps raid size at 200 runs per point and plots the asymmetry curve: how defense outcomes hold as the raid grows past the ready rack. To be precise about what it is: a presentation model with synthetic tracks and abstract probabilities, built to evaluate coordination logic, not vehicle guidance or fire control. That is the honest scope, and it is the scope where the architecture decisions get made.
What the auction shows when comms degrade
Three patterns hold across runs. First, the auction and the greedy planner converge when every interceptor is identical and every track is equal. Heterogeneity is what separates them: mixed interceptor types, mixed altitudes, and a defended asset make greedy assignment waste heavy interceptors on easy tracks and leave leakers for the expensive layers. Second, the auction degrades gracefully. Moving from permissive to denied comms costs intercept efficiency, but the fleet keeps fighting on pre-briefed policy and local scoring. The centralized baseline does not degrade. It stops. Third, policy matters more as comms get worse: under denied conditions, the policy the operator set before launch is the command intent, executed locally by every node.
That last point is where the autonomy policy debate lands in practice. DODD 3000.09 requires 'appropriate levels of human judgment over the use of force.' At Eurosatory, one analysis put the tension plainly: with hundreds of incoming targets, the human in the loop is the bottleneck, while French officials countered that 'we will make sure that we still have a man in the loop.' Both are right, and the resolution is architectural: human judgment sets the engagement policy, the approval gates, and the abort authority before and during the fight. The per-track assignment math runs at machine speed on the nodes. Judgment up front, arithmetic at the edge.
The same OS underneath
A CSIS analysis argues that software, not drones, will decide the next war. The airframes are becoming commodities. The Army's Low-Cost Interceptor program wants government-owned designs any manufacturer can build. What will differentiate defenses is the coordination layer: who assigns, how it survives jamming, and how it proves to a commander what it did and why.
That layer is not new territory for us. Distributed assignment over a lossy mesh, local inference under a compute policy, evidence capture with provenance, and mission-scoped destruction are the same primitives EdgeLance runs for ground ISR today. An interceptor is a tactical node with a motor. The drone mesh war in Ukraine already proved the two problems converge. If you want to see the simulator run your scenario, get in touch. It runs on a laptop, like everything else we build.