Agent-native
Agent-native describes software whose main interface is an AI agent, not a screen: you work in a conversation and the product holds the data behind it.
Three shapes get called agent-native, and they are worth separating because only one of them changes how the software is used. First: an application with an assistant added to it, where the interface is still the application's own screens and the assistant is a feature inside them. Second: an application that exposes a connector, usually an MCP server, so an outside agent can call it, while the application remains the place the work is done. Third: an application whose interface is the agent, with no screen the user is expected to open daily, and whose job is to hold the state and capabilities the agent reaches for.
The third shape is defined by what it does not ask for. Adopting software normally costs a habit, because something has to be opened, checked and remembered, and most tools that fail in the consumer market fail there rather than on capability. A product living inside an assistant somebody already opens inherits an existing habit instead of requesting a new one.
The constraint that comes with it is real and worth stating. Every function has to be expressible as a capability an agent can discover and call, an agent that cannot see a capability behaves as though it does not exist, and anything that genuinely needs a screen has to be worth the trip out of the conversation. Building in this shape is a set of trade-offs, not a badge.
Nothing about the term implies the agent is unsupervised. A carefully built agent-native product decides explicitly which operations an agent may complete alone and which require a person, and the second set is usually the ones that write.
Why it matters
The interface layer decides habits, and habits decide distribution. Software that assumes the user will open it competes for attention against everything else on the machine. Software that lives inside an assistant competes for something narrower: being what the assistant reaches for. The shape also determines the cost structure, because a product that runs no model of its own pays for no inference and gets better each time the frontier does, on somebody else's research budget.