What 'agent-native' actually means
Bolting an MCP endpoint onto a dashboard-first product does not make it agent-native. It means every mutation is a typed, idempotent, approvable operation.
"Agent-native" is on its way to becoming the most abused word in developer tooling. A dashboard with an MCP endpoint stapled to the side is not agent-native; it is a dashboard with a new way to click the buttons.
Agent-native is a property of the control surface, not a feature you add.
The four properties
Every mutation in Scribase — human or agent — has:
- A CLI flag and an OpenAPI operation. The same binary serves people and programs. There is no "API for agents" that drifts from the real one.
- An idempotency key. An agent that retries does not double-create. This is the difference between "mostly works" and "safe to automate."
- An operation ID. Every mutation returns one, so long-running work is observable and resumable instead of a fire-and-forget guess.
- An approval model. Destructive or production actions demand explicit confirmation. Autonomy without a gate is not a feature, it is an incident.
Why the gate matters most
The interesting failure mode of agents is not that they are dumb — it is that they are fast and confident. A control plane that lets an agent drop a table because it "seemed necessary" is a control plane that has not thought about agents at all. The approval model is where agent-native stops being a slogan and starts being trustworthy.
Read the CLI docs and the control API to see the shape of it.