Why on-device agents need their own operating layer
Most AI agents live in the cloud, and the cloud is forgiving. Compute is elastic, connectivity is assumed, and if something misbehaves you patch the server and move on. An on-device companion gives up all of those comforts the moment it is installed. It has to share a phone, kiosk or industrial box with everything else running there, keep working when the network does not, and protect the data it touches without a data centre's walls around it.
That is a genuinely different set of problems, and bolting a cloud-style runtime onto a device does not solve them. It is why LyboAI Edge companions do not run as loose apps. They run on Lybo OS — the on-device layer in the LyboAI stack whose whole job is to make an agent behave well on hardware it does not own.
What Lybo OS actually does
Think of Lybo OS as the companion's landlord. It hosts the agent that Edge Studio produced — the model, the tools it may call, and the guardrails it must respect — and it manages the practical realities of living on a device:
- Resource discipline. An agent should never be the reason a device feels slow. Lybo OS governs what the companion can use, so inference fits alongside the device's day job.
- Offline-first behaviour. The companion's core loop runs locally, so a dropped connection degrades features rather than killing the agent.
- Data boundaries. Conversations and business data are processed on the device by default. What leaves the device is a deliberate decision made at build time, not an accident of architecture.
The result is boring in the best way: the agent you tested in Edge Studio is the agent that runs in the field, inside limits the device owner can reason about.
Trust travels with the build
In Edge Studio, the last step before an agent ships is signing. That signature is not ceremony — it is the contract Lybo OS enforces. Before a companion runs, Lybo OS checks that the build it received is exactly the build that was signed: same model, same tools, same guardrails. If a package does not verify, it does not run.
The same rule applies to updates. A new version of a companion arrives as a new signed build, is verified on the device, and only then replaces the old one. That turns updating an agent in the field from a leap of faith into a routine, reversible operation — which matters when the device in question is a kiosk in a shopfront three hours from the nearest developer.
Real time, because the model is already there
Latency on the edge is not a tuning exercise; it is geometry. A cloud agent pays for a round trip on every turn — device to server, server to model, model back again. An on-device companion skips the trip entirely because the model is already sitting next to the conversation.
Lybo OS makes that proximity usable. It keeps the model warm, schedules inference so responses stay snappy even while the device does other work, and lets the companion react to local events — a scanned barcode, a sensor reading, a customer walking up to a kiosk — the moment they happen rather than after a network hop. For customer-facing agents, that difference is the line between a conversation and a loading spinner.
Where Lybo OS fits in the LyboAI stack
Each piece of the LyboAI stack answers one question. The VALUE-AI framework asks whether an agent is worth building at all. Edge Studio is where you build it — Canvas, projects, evals and signing. The Agents Platform runs cloud agents for the many businesses where cloud is the right call. Lybo OS answers the last question: once an agent has earned the right to live on a device, what does it live on?
We will be honest about where this sits: on-device is the newest frontier of our stack, and Lybo OS will keep evolving alongside LyboAI Edge as more companions ship. What will not change is the division of labour — Edge Studio decides what an agent is, Lybo OS makes sure that is exactly what runs.
If your business has a use case that should not depend on a network connection — or data that should not leave the building — that is precisely the shape of problem this layer exists for. Start with a conversation at lyboai.app.