Instrumenting Doorways

Every tool call is a doorway. Most instrumentation logs the room after the fact. The doorway is where the signal is.

The distinction sounds like a metaphor. It is an architecture.

What the doorway records

A tool call arrives with its arguments, its target, and its timing. The room, by contrast, is the state change left behind: the record deleted, the payment moved, the email sent. Logging the room means detecting the consequence. Logging the doorway means recording the intent, before the consequence exists.

The doorway record is small: what was called, with what arguments, against what authorization. It fits in a threshold budget that room-level telemetry never does. And the doorway is where intervention is still possible. Once the agent is in the room, the action is already underway.

Why the room loses

Consequence-level detection has three failures. It is late, because the consequence must exist before it can be logged. It is ambiguous, because the room shows the result but not the intent. And it is loud, because rooms are busy places and the anomaly is one transaction among millions.

The doorway is quiet. Agents make tool calls in patterns, and the patterns have shapes. An escape rehearses its exits. A fraud chain probes authorization boundaries before it uses them. The rehearsal is visible at the doorway and invisible in the room.

The implementation posture

Log the doorway by default, the room by exception. Instrument the tool-call boundary as close to the agent's intent as the stack allows. Rank the doorways by the four properties, spend the threshold budget there, and let the room-level systems do what they are good at: forensics.

The doorway is where the signal is. The room is where the story gets reconstructed. Instrument accordingly.