Review Continual Learning Improvements

Sikaru turns production misses, corrections, Experience trajectories, OpenInference-compatible traces, and eval results into reviewed improvements. The SDK exposes the review surface: list ready changes, inspect public diffs and evidence, approve to staging, and promote to production. This is Sikaru’s controlled continual learning loop for managed agents. The agent learns from failures and feedback, but behavior only changes through evidence-backed diffs, eval checks, and explicit release actions.

Lifecycle

Target	Purpose
`draft`	Editable behavior that is not serving traffic.
`staging`	Eval, regression checks, and team testing.
`production`	Live managed-agent behavior after promotion.

The SDK intentionally keeps promotion explicit. approve(..., target="staging") moves a ready draft to staging. promote(...) moves a staged change to production.

Review from Python

improvements = sikaru.improvements("support")
candidate = improvements.ready()[0]

diffs = improvements.diff(candidate.id)
evidence = improvements.evidence(candidate.id)

improvements.approve(candidate.id, target="staging")
improvements.promote(candidate.id)

Review from TypeScript

const improvements = sikaru.improvements("support");
const [candidate] = await improvements.ready();

const diffs = await improvements.diff(candidate.id);
const evidence = await improvements.evidence(candidate.id);

await improvements.approve(candidate.id, { target: "staging" });
await improvements.promote(candidate.id);

What diffs include

Diffs explain the visible behavior change Sikaru proposes. They are designed for product, engineering, support, and risk reviewers. Typical diff fields include:

The affected agent or behavior source.
A concise summary of the proposed change.
Before and after behavior where it is safe to show.
Linked checks or eval results.
The target release state.

What evidence includes

Evidence connects a proposal back to the reason it exists. Typical evidence includes:

Source traces and production examples.
Related user corrections or operator feedback.
Eval cases and pass/fail summaries.
Regression checks.
Reviewer decisions and release history.

Improvement responses are sanitized. They do not expose private prompts, generated skills, eval internals, optimizer state, model references, adapters, weights, or training corpora.

Endpoint mapping

SDK surface	Backing route family
`Agent.run`, `respond`, `stream`, `resume`	Managed run, event, recover, and tool-result routes.
`Experience` and `traces`	`POST /v1/trace-streams`.
`Capture.submit()`	Static eval references and improvement objectives.
`Improvement`	Project changeset routes.

Review checklist

Before promoting a change, confirm:

The proposed change addresses a real production miss or approved objective.
Evidence links back to concrete traces, corrections, or eval failures.
Regression checks cover the behavior you do not want to break.
The public diff is understandable to the people accountable for the agent.
Staging behavior looks correct before production promotion.

​Lifecycle

​Review from Python

​Review from TypeScript

​What diffs include

​What evidence includes

​Endpoint mapping

​Review checklist

Lifecycle

Review from Python

Review from TypeScript

What diffs include

What evidence includes

Endpoint mapping

Review checklist