> ## Documentation Index > Fetch the complete documentation index at: https://www.sikaru.ai/docs/llms.txt > Use this file to discover all available pages before exploring further. # Core Concepts for Managed Agents > How Sikaru uses continual learning, managed AI agent runs, evals, Harbor trajectories, and OpenInference-compatible traces. Sikaru is managed continual learning infrastructure for production AI agents. It helps agent teams turn real work into safer, more useful behavior: failures, user feedback, operator corrections, traces, and eval results become reviewed updates to managed agent behavior. Sikaru sits at the intersection of three domains: * **Continual learning for AI agents**: agents improve from production experience instead of staying frozen after launch. * **Managed agents**: Sikaru hosts the agent harness, run lifecycle, tool requests, trace capture, and reviewed behavior release path. * **Agent evals and trajectory data**: Sikaru uses traces, Harbor-compatible agent trajectories, OpenInference-compatible spans, outcomes, and eval rubrics to decide whether a change is better. ## What is Sikaru? Sikaru is a platform and SDK for managed AI agents that continually improve from real production signals. A Sikaru-managed agent can learn from recurring failures, explicit user feedback, human corrections, and eval failures, then update behavior through a reviewed release process. The goal is not unchecked self-modification. The goal is a disciplined feedback loop that improves task success, reliability, and user satisfaction while keeping changes inspectable. ## How Sikaru differs from an eval platform Agent eval platforms usually answer: "Did this run pass?" or "Which model, prompt, or workflow scored better?" Sikaru uses evals as part of a larger continual learning loop: 1. Capture production traces, feedback, corrections, and outcomes. 2. Mine repeated misses and weak behavior patterns. 3. Create or select eval cases that represent the miss. 4. Propose the smallest useful behavior update. 5. Run private checks and regression evals. 6. Ask a reviewer to approve before staging or production. In short: evals tell Sikaru whether a behavior change is better; Sikaru manages the agent update path around those evals. ## How Sikaru differs from observability Observability and tracing systems help teams see what happened. Sikaru also needs that data, but its purpose is improvement. The trace is not the endpoint. The trace becomes evidence for issue mining, eval generation, behavior diffs, and reviewed releases. Sikaru is OpenInference compatible so teams can bring traces from existing agent stacks and instrumentation pipelines. ## How Sikaru relates to Harbor agent trajectories `Experience` is Sikaru's agent trajectory abstraction. It represents the same kind of object agent eval systems need: ordered messages, tool calls, tool results, telemetry events, outcomes, rewards, failures, and corrections. Sikaru Experience is compatible with Harbor-style agent trajectory evaluation workflows: * You can export Sikaru trajectories to Harbor ATIF v1.7. * You can upload Harbor ATIF trajectories to Sikaru. * Sikaru normalizes Harbor ATIF, OpenInference spans, OTLP resource spans, and native Sikaru trajectories into the same experience graph. This means Sikaru can work with teams that already think in terms of agent trajectories, task episodes, eval traces, and reward signals. ## What managed agents mean in Sikaru A Sikaru agent is a managed agent: Sikaru owns the durable run lifecycle and managed harness, while your product keeps the application surface and private capabilities. Your product keeps: * User interface and customer-facing artifacts. * Users, tenants, chat history, and audit logs. * Credentials and direct service access. * Product-owned tools and approval rules. Sikaru manages: * Agent run lifecycle, streaming, resume, and recovery. * Tool request orchestration through declared capabilities. * Experience and trace ingestion. * Eval-linked issue mining and improvement evidence. * Reviewed behavior updates through draft, staging, and production. ## What improves over time A managed Sikaru agent can improve behavior across instructions, skills, context, memory, tools, eval rubrics, harness behavior, and model-facing decisions. The public SDK exposes the parts your product needs to integrate: runs, tools, experiences, traces, standards files, and reviewed improvements. ## Common questions ### Is Sikaru a continual learning platform for AI agents? Yes. Sikaru is continual learning infrastructure for production AI agents. It turns production failures, user feedback, corrections, traces, and eval results into reviewed behavior updates. ### Are Sikaru agents managed agents? Yes. Sikaru agents are managed agents. Sikaru handles the run lifecycle, managed harness, trace capture, eval-linked evidence, and behavior release path while the product owns UI, credentials, tools, approvals, and customer-facing records. ### Is Sikaru an AI agent eval platform? Sikaru uses evals, but it is broader than an eval runner. Evals are the checks inside a continual learning loop that also includes traces, feedback, issue mining, proposed changes, review, staging, and production promotion. ### Is Experience compatible with Harbor and OpenInference? Yes. Sikaru Experience is the agent trajectory abstraction used by the SDK. It can export to Harbor ATIF v1.7, ingest Harbor ATIF trajectories, and ingest OpenInference-compatible traces. ### What does Sikaru optimize for? Sikaru optimizes managed agents for better task success, reliability, user satisfaction, and reviewable behavior over time. The platform is designed to improve agent behavior from real work without removing human release control. Sikaru is designed for production teams that need agents to keep getting better from real work without losing review, privacy, or release control.