Automatic improvements
Automatic improvements is a bounded reliability loop. It watches the audit history for repeated failures, asks the configured model for a plain-language assessment, and may apply only predefined reliability tuning.
It is enabled by default in v0.4.0-beta.3. It is not one of the seven disabled-by-default Labs.
What it can change
Section titled “What it can change”The controller can select only allowlisted runtime reliability adjustments. For example, it may make a bounded retry adjustment after repeated transient model-provider failures.
Every applied change captures the previous configuration under the local improvements store. Application, failure, and rollback are recorded.
What it cannot change
Section titled “What it cannot change”Automatic improvements cannot:
- disable safeguards or approvals;
- expand network domains;
- grant permissions or capabilities;
- install extensions, agents, or skills;
- change credentials;
- edit source code;
- invent a new configuration action from model output.
The model can improve a proposal’s title, explanation, and priority. Code—not model text—decides whether an allowlisted action exists.
Inspect the current mode
Section titled “Inspect the current mode”odinn config self-improvement showodinn improve listRun a check immediately:
odinn improve learn --limit 1000Roll back an applied adjustment:
odinn improve rollback --improvement <improvement-id>Disable or re-enable
Section titled “Disable or re-enable”odinn config self-improvement set --enabled false --mode disabledodinn config self-improvement set --enabled true --mode autoNo embedded model is currently bundled with Ódinn. The assessment uses the configured provider or local model. Without a connected model, the loop can wait for a provider rather than silently substituting one.
Review Privacy to understand what a configured cloud model receives.