The main difference is what the server is trusted to do and what it's able to run. Architecture and Security go into the full design.
Device identity
The device's private key is generated on the device and never sent anywhere. During provisioning, which uses a short-lived single-use token, the server signs its CSR. Server to agent traffic is mutual TLS, and revocation is checked against the server's records on every request instead of trusting the handshake.
Human authentication
WebAuthn/FIDO2 only. There's no password sign-in in normal operation.
Typed jobs
The list of actions is short and fixed: system.status, systemd.list, systemd.status, apt.list_updates and systemd.restart. Arbitrary shell execution isn't on the list and won't be. Jobs are signed by the server and carry a nonce and an expiry so they can't be replayed. systemd.restart also needs step-up, and is only created and dispatched for a TRUSTED device. Jobs are submitted through the server API for now.
Privilege separation
lg-agent runs unprivileged. Anything that needs root goes to a separate, tightly scoped lg-executor process over a local socket, and the executor checks each job's signature again before it runs. A compromised agent can't just run whatever it wants as root.
Step-up authentication
A fresh WebAuthn check, enforced by the server, is needed for device revocation, attestation re-baselining, adding a security key and the privileged systemd.restart job. Each SSH certificate request gets its own single-use check, tied to that request. Read-only jobs only need a valid session.
Audit trail
Sign-ins, revocation, the job lifecycle and SSH certificate issuance are written to an append-only audit log on the server. There's no interface for browsing it yet.
Proxmox correlation
Read-only, against one Proxmox cluster: VM and container inventory and power state, matched against agent connectivity so you can see when what Proxmox says and what the agent says don't agree. No start, stop, reboot or other lifecycle actions.