Two Loops of Evolution

Online loop only executes tasks and logs evidence – does not modify the live agent.
Offline loop aggregates trajectories, diagnoses root causes, creates candidate revisions, and publishes a new version after passing validation thresholds.

Loops are connected by a versioned experience base and evaluation suites.

Separating logging from systematization prevents immediate modification due to fluke success, network glitch, or adversarial input and enables larger batches with cheaper models.

Voyager: automated curriculum + skill library + environment vetting – all three are necessary.