Skip to content

The closed loop

A closed loop in the AI-native world is the same shape as a control loop in industrial systems: setpoint, plant, sensor, controller, actuator, eval.

Loop termAhamie equivalent
SetpointThe spec — what success looks like
PlantThe connector + tools layer (the world the agent acts on)
SensorRunOutcome — populated by systems the agent does not write to
ControllerThe agent loop
ActuatorThe action set — agent.run, app.invoke, gateway.send
EvaldefineSuite + hidden-golden

Without a sensor that the agent cannot influence, you can’t separate good agent behavior from agent-self-confirmation. The agent will tell you it succeeded; the question is whether anyone else agrees.

The hidden-golden partition + sensor isolation are the two primitives that keep the loop honest.

Mastra owns the controller (Agent), the actuator wiring (tool call), and the plant adapters (model gateway, sandbox, FS). Ahamie owns the sensor, the eval, the trust boundary on the plant, and the outer loop that revises the spec when the eval fails. That’s the whole bet.