Edge and remote sites
Inference and apps that keep working when the link to the core does not.
What it is #
Remote sites break the usual assumption that the network is there. An Agentycs edge is a self-contained inference site: it hosts its own workers, serves requests from its own accelerators, mirrors the model artefacts it needs locally, and enforces exactly the same tenant policy the core would.
The link between edge and core is an authenticated, signed HTTP/3 bridge that expects loss and variable latency, rather than a local network protocol pretending to stretch across a continent. Control and dispatch flow over it, token streams recover across transient loss instead of dropping the generation, and if it goes away entirely the site keeps serving what it has loaded and reconciles its state when it comes back. What it tolerates on your particular link is a sizing question we would rather answer against your conditions than in a brochure.
Be precise about what that buys you. Inference has no single dependency on the core: a site serves from its own accelerators and its own mirrored weights, so losing the link degrades the fleet view rather than the answer. Identity, authorisation and the audit path are a different matter, and so is the first cold start of a model a site has never held. Those are bounded by policy you set, not by physics, and they are worth deciding before you need them.
Who it is for
- Your work happens where connectivity is intermittent, expensive or contested.
- Latency to a central region is unacceptable for the decision being made.
- Data is generated at the site and should be reasoned about there.
- You need one fleet view of many sites without depending on all of them being reachable.
A site serves the models it holds. Everything else about the edge follows from that one sentence.
What you get #
The whole platform runs here. These are the parts this pattern changes the shape of.
A real edge runtime #
Each site hosts model workers and serves locally, with the same authorisation and tenant-policy enforcement as the core rather than a relaxed variant.
Weights already local #
The artefact mirror seeds, pins and holds model weights at the site, so models load and serve even while the bridge is down.
Autonomy that reconciles #
A site keeps serving its loaded models if the core is unreachable, and the central view reconciles what it was actually running once the link returns.
Fleet-wide placement #
The core publishes desired model state out to sites and reads back what they are running, so a hundred sites are one inventory rather than a hundred conversations.
The operating model #
Nobody should discover who owns a failure while it is happening. This split is explicit before anything is signed, and it is the part of a deployment decision that outlives the architecture.
- The core team operates
- Fleet-wide model state, rollout waves, revocation, and the single view of every site's accelerators.
- The site operates
- Local hardware, power, environment and physical access, plus whatever local workflows the site owns.
- The bridge handles
- Authenticated control and dispatch, stream recovery across loss, and state reconciliation on reconnect.
- You keep
- The policy that decides what a disconnected site may still do, and how long stale state remains valid.