Skip to main content

  1. Agentycs
  2. Industries
  3. Defence and intelligence

Defence and intelligence

Capability that only works when data crosses a boundary is capability you cannot use.

The operating reality #

Defence and intelligence work is organised around boundaries: classification, releasability, need-to-know, and a clear chain of authority for what may be acted upon. An AI service that requires data to leave the enclave fails the first test, however good its answers are.

There is a strategic version of the same test. A national capability rented from a foreign commercial platform is a capability that another government can restrict, reprice or reprioritise, and it can do so at exactly the moment the relationship between two states has changed. Sovereignty here is not a preference about hosting. It is the difference between an option you hold and an option somebody else holds over you.

The second constraint is physical. The places where an answer matters most — a forward site, a ship, a vehicle, a deployed headquarters — are the places with the least bandwidth and the least reliable link. Planning staff do not build a capability into an order of battle unless they know what it does on the day the link is not there.

So the design question is not how clever the model is. It is what each site holds before it deploys, what it can still do on its own, what a headquarters can and cannot see while it is out of contact, and what reconciles when it comes back.

A capability that degrades to nothing when the satellite drops is a capability nobody will plan around.

What that requires of the platform #

  • The whole stack deploys inside the enclave, air-gapped where required, with no runtime dependency on an external model host.
  • Inference runs at the site on hardware sized for it, down to a backpack-sized 6-litre micro server drawing 400W and carrying a full 96GB GPU, and keeps serving its loaded models when the link to the core is gone.
  • Retrieval enforces the source system's permissions during the query, so a cleared answer stays cleared and an uncleared source is never surfaced.
  • Tenant separation is derived server-side from validated identity, never from anything a client can assert.
  • Every data access and privileged action lands in a canonical, redacted audit ledger correlated to the identity behind it.
  • Sites and services authenticate each other across a zero-trust overlay with its own internal naming, over networks you do not own.

What the site holds, and what the bearer carries #

The useful question about a deployed capability is not how it is drawn. It is what is on the ground at the site, what only exists back at the core, and which of the two the link is standing between.

A sovereign core, a bearer, and a deployed site A topology in three parts. On the left, the sovereign core holds the model hub of record, identity and policy, a single view of every site, and the audit archive. On the right, a deployed site holds mirrored model weights already on disk, model workers serving requests locally, the same tenant policy as the core, and its own audit records. Between them runs the bearer, carrying control and dispatch outwards and state reconciliation back. The bearer is marked as the only part of the topology that can be absent.
The link What the site does What the core sees
Link up Serves requests locally from mirrored weights, under the same tenant policy as the core, and reports what it is running. Publishes the model state it wants at each site, and reconciles what each site is actually running into one view.
Link down Keeps serving the models it has already loaded, keeps enforcing the same policy, and keeps its own audit records. Nothing new can be mirrored to it. Holds the site's last reported state. It cannot see current activity and cannot push a new model out until the link returns.
Link restored Reconciles its state and collects whatever was published while it was out of contact. An in-flight generation recovers rather than being dropped. Returns to a current view of the site without anyone needing a session at the site itself.

Scope a site, not a pilot

Bring the site, the classification it sits at and the longest period it has to keep working alone. We will show you what runs there, what it holds and what an operator sees while it is out of contact.

Agentycs is pre-revenue and signing its first deals with its first customers. We name no deployments and no users on a public website, and we are not going to invent one. Ask us directly and we will tell you exactly where we are.