A copilot can whisper answers. It cannot sketch and talk at once.

how the system behaves, not a benchmark

A fluent answer is not a demonstrated one.

Sounding right in the moment is exactly what a copilot is good at. Thinking on your feet without one is a different thing, and it is the thing the interview is built to surface.

The layer underneath: AgentR Guard.

Everything above is about how the interview itself resists a copilot. Guard is about the machine it happens on. A desktop app for macOS and Windows, running for the length of the session, watching the environment rather than the answers.

• the machine
  Processes, windows and screen capture, read at OS level

  An overlay built to hide from screen sharing is still a process on the machine, still drawing a window, still reading the display. That is the layer Guard sits at, which is why a tool designed to be invisible to a video call is not invisible to it.

• the network
  Where the machine is talking during the session

  An answer arriving from somewhere is traffic before it is ever text on a screen.

• camera and mic
  Who is in the room, and who is speaking

  A second face, a phone held below the desk, a voice that is not the candidate’s.

• the room
  It can hold the session shut until the machine is clean

  When something is found the interview stops, rather than carrying on compromised, and reopens once it clears. That is a pause and a marker for a person to review. It closes a room; it never closes the door on a candidate.

It works when a person is asking the questions too

Guard is not tied to the AgentR interview. Run it alongside a human interview on whatever call software you already use, and the same layer is watching for the same overlay tools. The conversation stays yours. Only the machine is checked.

What a session gets read as.

Not every mechanism that protects an interview is published, and that is deliberate. What is fixed is how the reads are defined, and that no single soft signal reaches one on its own.

• clean
  No anomalies across any layer

  The default read for the overwhelming majority of interviews.

• minor flag
  One soft signal, on its own

  A pause, a glance away. Logged, and not enough by itself to mean anything.

• elevated concern
  Multiple signals corroborating each other

  Several soft signals lining up in the same direction, worth a look before you decide.

• confirmed anomaly
  Hard evidence, not an inference

  Something like clipboard content matching text found elsewhere. Rare, and clearly flagged when it happens.

What it will not do.

• It will not flag or fail a candidate off a single soft signal alone. Corroboration across layers is required.
• It will not show the candidate which mechanisms are active, or show recruiters a raw signal-by-signal log. Only a categorical read.
• It will not replace the recruiter's judgement. It surfaces integrity concerns; a human makes the call.
• It will not penalise normal nervousness or an unusual setup as if it were cheating.