Platform class sets the physics. Small UAS under 2kg operate on milliwatt SWaP budgets and sub-20ms perception loops; an MQ-9-class Group 5 platform has orders of magnitude more compute but must still hold a sub-100ms loop under SATCOM-degraded conditions. A single stack rarely spans small-UAS and Group 5 without material subsystem divergence — the most common architectural failure is designing the stack against one class and retrofitting the others.

At 10 m/s (typical UGV speed), a 50ms perception-action latency means the vehicle has already moved 50cm before acting on the last sensor reading. The budget is set by the physics of platform speed and stopping distance, not by engineering preference — and once set, every millisecond spent on network egress is a millisecond unavailable for perception, planning, or safety arbitration. Budgets must be chosen first and architecture chosen to fit them.

DoD Directive 3000.09, reissued Jan 25, 2023, imposes specific review and approval paths on lethal autonomous weapons systems. Semi-autonomous and autonomous systems that select and engage targets without human action require senior-leader review (USD(R&E), USD(P), VCJCS) before formal development and again before fielding. Non-weapon autonomous systems (autonomous navigation, ISR collection, logistics) do not carry the same engagement-approval requirement, but still inherit DoD AI Ethics Principles and LOAC obligations.

SWaP is the hard constraint that kills more autonomy programs than accuracy does. Model architectures that hit target accuracy at server-class power budgets routinely miss by 5–10x at embedded envelopes, and the gap is not closed by quantization alone. SWaP must be budgeted before model selection — not discovered after training. The Skydio X10D, Anduril Bolt, and Shield AI V-BAT all ship specifically because their autonomy stacks were designed to the SWaP envelope, not retrofitted into it.

Replicator and the DoD Joint Warfighting Concept both assume comms-degraded and GPS-denied operation as the baseline, not the exception. A stack whose safety case only holds with continuous link and GPS is not a fielded autonomous system — it is a remotely piloted system. Environmental scope must be stated explicitly so perception training data, SLAM fallback, and degraded-mode behaviors can be validated against it.

DARPA OFFSET's final field experiment at Fort Campbell in 2021 demonstrated 300+ combined physical and virtual autonomous agents operating simultaneously in urban terrain — this remains the largest coordinated swarm experiment DoD has publicly validated. Swarm scale changes the architecture: a pair can run on a centralized controller; a 300-agent swarm requires mesh coordination with no central dependency because any central node is a single point of failure and a communications magnet.

Autonomy software that directs weapons employment or enables military-critical capabilities generally falls under USML Category XI (military electronics) or Category VIII (aircraft) on the U.S. Munitions List, making it ITAR-controlled. ITAR restricts who can touch source, training data, and model weights — misclassifying early leads to painful contractor and cloud-vendor rework. The classification must be made before vendor selection, not after.