Skip to main content
The Atlas SDK orchestrator wraps every request in an adaptive dual-agent loop. The runtime triages the task, probes capability, and then pairs your agent (the student) with a verifying teacher that applies the right level of oversight before results ship.
Looking for the personas behind each role? Read the Student & Verifying Teacher Roles guide after you understand the controller shown here.

​
Runtime Diagram

Adaptive runtime flow showing triage, probe, and lane routing into auto, paired, coach, and escalate before telemetry and rewards.

The runtime triages tasks, probes capability, and routes them into the correct supervision lane before the student + verifying teacher execute.

​
Flow Breakdown

1

Triage the request

A triage adapter builds a triage dossier from the incoming task, prior personas, and recent telemetry. Custom adapters can enrich the dossier with business-specific metadata.
2

Probe capability

A lightweight probe model scores confidence and surfaces evidence. The orchestrator stores the score, evidence, and lane suggestion in adaptive_summary.
3

Select supervision lane

Based on probe confidence and certification requirements, the orchestrator picks auto, paired, coach, or escalate. Manual overrides and mode pins come from adaptive_teaching.
4

Execute the dual-agent loop

The student plans and executes, while the verifying teacher reviews, validates, and issues guidance according to the active lane. Retries only fire in lanes that permit them.
5

Score and persist

RIM judges score the outcome, certification verdicts are recorded, and telemetry flows to the console or storage backends. Export these traces with arc-atlas for offline training.
⌘I