Why building for the edge needs its own workshop
Shipping an AI agent to the cloud is forgiving. If a prompt misbehaves, you patch the server and every user gets the fix in seconds. Shipping an agent to a device is a different discipline: the model, the tools and the behaviour all travel together, run without a guaranteed connection, and sit inside someone's phone, kiosk or industrial box where a bad release is genuinely hard to claw back.
That is the problem Edge Studio exists to solve. It is the workbench behind LyboAI Edge, our on-device AI companions, and it treats the journey from idea to installed agent as one continuous pipeline rather than a handful of disconnected tools. This post walks through that pipeline the way a team actually uses it: Canvas to project, project to evals, evals to a signed build running on Lybo OS.
Canvas: where an agent takes shape
Every agent starts on the Canvas. It is the visual surface in Edge Studio where you lay out what the agent is: its role and tone, the tools it may call, the data sources it can reach, and the guardrails it must respect. Because the Canvas is visual, the people who understand the business — not just the people who write code — can see and shape the agent's behaviour before anything ships.
From the Canvas, work is organised into projects. A project holds the agent definition alongside everything that belongs to it: prompt versions, connected tools, test conversations and eval suites. That keeps a team's iterations in one place, so "which version are we actually shipping?" has a boring, reliable answer.
Evals: proof before promotion
The step most teams skip is the one Edge Studio makes routine: evals. An eval suite is a set of scenarios the agent must handle well — the common requests, the awkward edge cases, the things it must refuse. Each candidate build runs against the suite, and the results tell you in plain terms whether this version behaves better or worse than the last one.
This matters doubly on-device. In the cloud you can hotfix your way out of a regression; on a fleet of devices you want the regression caught before the build leaves the studio. The loop is deliberately simple: edit on the Canvas, run the suite, read the results, repeat until the build clears the bar you set.
Signing: the step that makes a build trustworthy
Once a build clears its evals, Edge Studio signs it. Signing produces a sealed bundle — the agent definition, its configuration and its eval record — with a cryptographic signature that says: this exact build came from this workspace and has not been altered since. The device checks that signature before it runs anything.
For anyone operating agents in the field, this is the difference between "we think the kiosks run version 12" and knowing they do. It also gives you a clean audit trail: every signed build traces back to a project, its eval results and the person who promoted it. No mystery builds, no drift between what was tested and what is running.
On the device: Lybo OS and the Edge runtime
The signed bundle lands on hardware running Lybo OS, the layer that hosts LyboAI Edge companions. Lybo OS verifies the signature, loads the agent, and gives it a controlled runtime: local inference where the hardware allows it, managed access to the tools the bundle declares, and sensible behaviour when the network drops — because on the edge, it will.
Running on-device is not just an infrastructure preference. It means conversations and data can stay local by default, responses do not depend on a round trip to a data centre, and the agent keeps working in the environments where connectivity is worst and the work is most real — shop floors, vehicles, clinics, counters.
Where VALUE-AI fits, and where to start
The pipeline tells you how to ship an agent; our VALUE-AI framework is how we decide whether and what to ship — starting from the business outcome and working back to the agent, not the other way round. Edge Studio is the place those decisions become running software, with evals and signing keeping the honesty in the loop.
If you are weighing up an on-device agent for your own operation, start small: one workflow, one eval suite, one signed build. Explore Edge Studio and the LyboAI Edge companions at lyboai.app — and if you want a walkthrough of the Canvas with your use case on it, that is exactly what we like doing.