Skip to main content
Settings → Loop configures the self-improvement loop: where it runs, which of its agents are on, how far automation goes, and how much it may spend.

​
Runtime

  • Cloud: the loop runs on Nora’s infrastructure.
  • On-premises: the loop’s own model calls route to an endpoint you registered under Internal resources.

​
Automation agents

Toggle each of the loop’s agents individually:
  • Verification: verifies runs automatically.
  • Failure detection: detects failing patterns and raises signals.
  • Improvement agent: proposes fixes for confirmed problems.
  • Evaluation agent: scores candidates against your evaluation criteria.
Verification lenses holds the Causal check (auto) toggle. Once on, it runs automatically.

​
Autonomy level

How much the loop does without a human in the loop. Level 3 and Level 4 are selectable today; Level 4.5 and Level 5 are marked upcoming. Higher levels hand more of the detect → propose → validate cycle to the loop.

​
Budgets

  • Max spend per run: a single loop run stops at this spend (default $10).
Simulation & optimization budgets bound sweep work:
  • Sweep budget: total spend per sweep (default $10).
  • Timeout: per sweep (default 10 minutes).
  • Cases per candidate: how many dataset cases each candidate is evaluated on (default 6).

Optimization

What the loop’s proposals look like.

Simulation

The sweeps these budgets bound.

Usage

Loop model tokens are metered on the verify & improve axis.