priorauth.indocs

Agent control plane

Five agents can read charts, ask payers about members, write into clinical records and transmit authorizations. Somebody has to be able to answer, at 2am, on a phone: what are they allowed to do, what have they done, and how do I stop one right now.

#The asymmetry

Stopping is frictionless. Starting is not.
No approval, no proposal cycle, no second signature to pause an agent. An incident does not wait for a quorum, and a control that is slow in the dangerous direction gets routed around by whoever is under pressure. Restarting requires a written reason, is attributed, and returns a quarantined agent at a reduced ceiling.

#States

StateMeaning
activeRunning normally.
degradedRunning, but capped below its configured autonomy - usually a failing or stale eval suite.
pausedStopped by a person.
quarantinedStopped by a guardrail. Nobody was watching and something tripped a wire.

The distinction between paused and quarantined matters: an operator knows why they stopped something. A guardrail stop means the condition arose unobserved.

#The fleet switch

One lever halts every agent at every practice, immediately. A fleet stop is reported as fleet_stopped, not as a permissions error - it is a platform-level halt and it should not read like a scope problem.

#Autonomy is licensed by evals

An agent may act above L1 only while a passing eval suite says it should, and the licence expires after fourteen days. A lapsed agent drops to proposing rather than stopping - losing the ability to act is not the same as losing the ability to help.

ConditionEffect
No eval run has ever happenedCapped at L1, reason no_eval_run
The suite is failingMarked degraded, capped at L1
The last pass is older than 14 daysCapped at L1, reason eval_stale
A never-event occurredQuarantined immediately, with nobody watching

#Per-practice control

A practice can decline an agent the rest of the fleet runs, without contacting anyone. Other practices are unaffected.

#Reach

The first question in an incident is not what an agent was configured to do - it is what it could have touched. The console shows, per agent: whether it reads PHI, whether it can write charts, whether it transmits to payers, how many of its tools are gated, and how many ungated writes it has. That last number should always be zero.