Glove is not an agent builder. It is a deterministic state machine that sits at the centre of your operation, and assembles ephemeral agents at runtime from the state it holds.
{{ e.body }}
Deterministic doesn't mean rigid. Every turn, Glove evaluates intent — and rules respond to real business variables. The same process behaves differently for a €500 contract and a €50,000 one.
Draw the states, the fields each one collects and where every outcome routes. The engine turns that into the agent it assembles at runtime.
A form, a message, a call, or a spreadsheet sync — the engine receives it, processes it, and coordinates the right response.
A thread started on WhatsApp continues on voice without losing context, because the channel is a transport — not a memory.
Any MCP server becomes an action the engine can take at the right state — calendars, CRM, ERP, search, external models, your own API.
Glove runs on its own. It holds what each process needs to execute and to learn from as it runs — its own execution record, not a copy of your CRM. Two records, not one: Glove keeps the execution record, your CRM keeps the business record. Connect them and every outcome is written there too, so your existing systems keep doing their job.
The flow is deterministic; the wording is generative. The model produces the language of each turn, but it never decides which state comes next — the engine does, from the process you defined. The probabilistic layer lives inside bounded actions, never in routing.
No. FlowBuilder is visual: you define states, transitions and the actions available at each step. Connecting an MCP server or a channel is configuration, not custom code.
Because a standing agent carries its own drifting memory, and two people end up with two different behaviours from the same process. Assembling one per turn from the process definition and the current state removes that: the agent is discarded, the state persists, and the ten-thousandth turn behaves like the first.