A Relay

A Relay is the mechanism by which a Post transfers a task or decision to a Station. It is not a failure. It is the system working as designed. A Post that never Relays is not more capable — it is more dangerous, because it means the agent is attempting to handle situations it was never scoped for.

The Relay is what makes the human-agentic model governable. Without it, you have two options: restrict the agent so heavily that it handles almost nothing, or give it broad autonomy and hope it makes good decisions. The Relay is the middle path — the agent operates autonomously within its scope and escalates the moment it reaches a boundary.

The name comes from relay racing: a clean handoff at a defined point, with both parties knowing exactly when the transfer happens and who holds the baton afterward. A Relay is not "the agent got confused." It is "the agent reached a condition where human judgment is required, and it transferred cleanly."

Escalation implies failure. Something went wrong and now it has to go up the chain. That framing creates incentives to avoid escalating. Managers who track "escalation rates" as a negative metric are training their teams — and their agents — to handle things they should not be handling in order to keep the number low.

A Relay is a normal, designed, expected handoff from a Post to a Station at a trigger point that was engineered before the Post was deployed. The word changes the culture around it. Relay volume is a diagnostic, not a failure metric. Too many Relays means the triggers are too sensitive or the scope is too narrow. Too few means the agent may be handling things outside its governed territory. Both extremes are informative.

A healthy Post Relays regularly. If your agent has not Relayed in a month, that is not a sign of capability. That is a signal to check whether the trigger conditions are tight enough.

A Relay is triggered by a condition defined in contracts.md. Common triggers include: confidence below a threshold, input that falls outside the training distribution, a request that touches out-of-scope territory, any irreversible action, and any situation where the agent detects ambiguity it cannot resolve.

The challenge with current AI models is that they often do not know they are uncertain. A language model will produce confident output even when hallucinating. This means Relay triggers cannot rely solely on the agent's self-assessment. They must also include external checks: output validation rules, schema conformance, human-in-the-loop sampling, and anomaly detection.

The most reliable Relay triggers are deterministic, not probabilistic. "If the input contains a legal question, Relay" is more reliable than "if the model's confidence is below 0.7, Relay" — because confidence scores are often miscalibrated. The Tripwire Doctrine says it directly: a Relay is a tripwire engineered by the Builder. If the Builder does not code the exact condition, the Post will not stop. If your escalation design relies on the agent recognizing its own uncertainty, you do not have escalation. You have hope.

A missing Relay is the second most dangerous failure mode in AI-native operations, after a Veer. When an agent encounters a situation it should escalate but does not, the resulting output is produced without the human judgment the situation required. The agent does not flag it. The human does not review it. The error enters the system as if it were normal output.

The fix is testing Relay triggers with the same rigor as testing the agent's primary function. testing.md should include known edge cases where the agent should Relay but might not. The weekly Bearing should include a review of Relay volume — both too many and too few.

Relays are processed at the Bearing. This is the integration point between the agent governance and the human cadence. A Post escalated something. What is the resolution? If the Post is Relaying repeatedly on the same trigger, is there a Groundwork update needed or is the trigger working correctly and the resolution is the point? The Bearing is where the Relay signal becomes a governance decision.