Sikaru SDK: Continual Learning for Managed Agents

The Sikaru SDK is a small set of server-side primitives for running managed agents and feeding production experience back into their behavior. Use it to start managed runs, expose product-owned tools, capture agent trajectories, upload OpenInference-compatible traces, connect eval evidence, and review behavior updates before they reach users. Sikaru is built for production agents that should improve from real work without losing release control. Failures, user feedback, corrections, outcomes, and eval results become evidence for safer and more useful managed-agent behavior.

Why use Sikaru

Sikaru gives agent teams a practical loop for continual improvement:

Run managed agents from your product server with durable streaming and resume.
Keep tools private by executing product capabilities inside your infrastructure.
Capture Experience trajectories that are compatible with Harbor-style agent trajectory workflows.
Bring existing traces through OpenInference-compatible spans, OTLP resource spans, or native Sikaru trajectories.
Use evals as release checks before managed behavior changes reach staging or production.
Review every update with public diffs, evidence, and explicit approve/promote actions.

Start here

Quickstart

Install the SDK, configure a project key, declare a tool, and run your first managed agent.

Core concepts

Learn how managed agents, evals, Harbor trajectories, and OpenInference-compatible traces fit together.

Choose a workflow

Managed agent runs

Use respond, stream runs manually, resume durable runs, and register local or provider tools.

Experience ingestion

Record Harbor-compatible agent trajectories, OpenInference spans, signals, outcomes, failures, and raw traces.

Standards files

Define managed behavior with instructions, skills, eval rubrics, and explicit trace uploads.

Review improvements

Inspect diffs and evidence, approve to staging, then promote reviewed behavior to production.

Core primitives

Primitive	Use it for
`Sikaru`	Project-scoped client configured with a server-side API key.
`Agent`	Managed agent entrypoint for a Sikaru-hosted harness.
`Run`	Durable execution you can stream, resume, and feed tool results into.
`Tool`	Product-owned local or brokered capability Sikaru may request during a run.
`Experience`	Sikaru’s agent trajectory abstraction: messages, tool calls, telemetry, outcomes, corrections, and rewards. It is Harbor ATIF and OpenInference compatible.
`Improvement`	Reviewable behavior change with public diffs, eval evidence, and release actions.

Continual learning loop

Sikaru does not let production agents rewrite themselves without review. It turns real-world evidence into proposed updates, checks those updates against evals and regressions, and keeps humans in the release loop.

Request flow

Keep Sikaru API keys on application servers only. Do not ship them to browsers, mobile clients, or customer-visible automation sandboxes.

Packages

Python
TypeScript

pip install sikaru-sdk

from sikaru_sdk import Sikaru, tool

npm install @sikaru/sdk

import { createSikaru, tool } from "@sikaru/sdk";

Authentication

All SDK calls use a project-scoped API key.

export SIKARU_API_KEY="sk_sikaru_..."
export SIKARU_PROJECT_ID="proj_..."
export SIKARU_API_URL="https://api.sikaru.ai"

The default API URL is https://api.sikaru.ai. You only need SIKARU_API_URL when targeting another environment.

Project keys are created after signing in to Sikaru and provisioning a project. The platform shows the key once during setup and supports rotation from project settings.

​Why use Sikaru

​Start here

Quickstart

Core concepts

​Choose a workflow

Managed agent runs

Experience ingestion

Standards files

Review improvements

​Core primitives

​Continual learning loop

​Request flow

​Packages

​Authentication

Why use Sikaru

Start here

Choose a workflow

Core primitives

Continual learning loop

Request flow

Packages

Authentication