Same node, different body
- date
- 2026-08-17
- reading
- 6 MIN
- author
- SKYSHIELD TECHNOLOGY
Every autonomy company eventually claims its stack is "platform-agnostic." The claim usually means something narrower: we have ported it twice, and the second port hurt less than the first. That is not platform agnosticism. That is scar tissue.
Real platform agnosticism is a decision you make before the first platform exists, and it shows up as three disciplines that are easy to state and expensive to keep.
The node never knows its body
The first discipline is a prohibition: no part of the decision-making core may name the platform it runs on. Not its type, not its dimensions, not its manufacturer, not a single hardcoded number that quietly encodes "this is an aircraft."
What the core is allowed to know is what the platform declares: what it can do, within what envelope, at what cost. The platform describes itself; the core plans within the description. The moment one magic number leaks — one assumption that endurance is measured in hours, or that motion happens in three dimensions rather than two — the abstraction has failed, and you will not find out until the port.
Intent goes down, never commands
The second discipline shapes the interfaces. If the layer above sends commands — go here, turn there — it must understand the body it is commanding, and the abstraction dies at the first vehicle that turns differently. So the layers exchange intent: objectives and constraints, decomposed downward, with each node deciding how inside the bounds it was given.
A search objective means one thing decomposed across a formation of aircraft and another thing decomposed across ground vehicles in a valley — and the layer that issued it does not need to know the difference. That indifference is the product.
The proof must be brutal
The third discipline is about honesty. A stack that has flown two platforms has proven it can fly those two platforms. If the architecture claims to be platform-agnostic, the test must be one no team can quietly tune for: platforms the core has never seen, generated in quantity, with envelopes chosen adversarially.
The test is not whether the system flies each one gracefully. It is whether the core needed zero changes to try. A port that requires touching the core is not a port; it is a rewrite wearing a port's clothes.
None of this is free. Keeping the core body-blind costs abstraction work up front, and the temptation to "just read the airframe config from here" arrives weekly. The reason to pay is at the bottom of every roadmap: the fourth platform, the fifth domain, the integration you have not imagined yet. Architecture is what you buy before you know you need it.