Skip to main content
Optimization is the machinery between “we have a problem” (a signal or cluster) and “the problem is fixed” (a shipped version). It proposes concrete changes, tests them, and hands them to a human for approval. You don’t have to use it — you can hand-fix everything through the normal Draft-and-Publish flow. But Optimization automates the ceremony: proposing, testing, and preparing changes for review.

​
The flow, in six steps

  1. Diagnose — pick a cluster (or signal). Nora locates the root cause.
  2. Propose — Nora generates 1-3 candidate fixes targeting the root cause.
  3. Pick a lever — you decide which part of the system to change: prompt, tool, retrieval, memory, etc.
  4. Simulate — every candidate runs against your dataset. Before/after numbers.
  5. Approve — pick the winner. Optional gates (human review, code review, testing).
  6. Deploy — the change applies to Draft; you publish (or auto-publish if configured).

​
Pages

Diagnose a failure

Root cause identification.

Improvement proposals

The candidate fixes and their diffs.

Which lever to tune

Prompt vs. tool vs. retrieval vs. memory vs. rules.

Approve & deploy

Human gates and rollout.

Improvement history

Every improvement, its result, its status.

​
Automatic vs. manual

The whole flow can run automatically for low-risk changes — you set the thresholds and Nora ships fixes without waking you up. Or it can require human approval at every step. Recommended: automatic for well-understood levers with strong sim results (small prompt tweaks, rule additions). Manual for anything model-swap, tool-config, or memory-structural. Configure per Flow: Flow settings → Improvement automation.

​
Reproducibility

Every improvement carries:
  • The cluster that seeded it.
  • The proposals that were considered.
  • The simulation results (target vs. baseline).
  • The approver.
  • The deploy timestamp.
Everything is reviewable months later. Great for compliance audits, incident review, and understanding why the Flow is the way it is.

​
Improvements and versions

Every deployed improvement creates a new Flow version with the improvement noted. Rollback rolls back the whole improvement atomically. You can see improvements attached to any version — “v18 was published to fix cluster #142 (billing hallucinations).”

​
When improvements aren’t automatic

Optimization proposes changes, but some fixes require things Optimization doesn’t do:
  • Adding a new document to knowledge — you upload.
  • Building a new tool — you build.
  • Fixing a broken upstream API — you fix.
For these, Optimization surfaces a root-cause note in Signals and stops. You take it from there.