Our own agent engine.

• Plans and queues the work.

  Every piece of agent work is a task in a durable queue, with retries, time limits and a clean cancel.

• Keeps memory separate.

  Each company has its own facts, rules and lessons. Agents read them before every task. Nothing crosses between companies.

• Picks the right model.

  Each task class has a route with fallbacks. Bulk public work goes to low-cost open models; personal data goes only to approved models.

• Holds the budget.

  Limits per task, per company per day and per plan per month, with spend recorded on every task.

• Checks every result.

  A task is done only when its result passes the checks for its kind. Failed work is retried or flagged, never billed.

• Stops at your rules.

  Every action that leaves the system is checked against consent, unsubscribes, daily limits and your approval level first.

• Tests before it ships.

  Prompt and model changes run against test sets and must do at least as well as the version before.

• One switch to stop.

  Pause everything, one company, or one kind of action.

We are not tied to one model company. Today bulk work runs on low-cost open models with fallbacks, and personal data runs on models approved for it. When a better or cheaper model arrives, we change a setting, not the product.

Our studio builds custom agents for clients on the same harness, with the same checks and budgets.