Build, run, and improve agents with Sikaru | Sikaru

Define what an agent should do, give it the tools it needs, and send it work. Sikaru runs the agent, keeps its execution state, and gives your application events, results, and evidence of what happened.

Use it for a support assistant that investigates cases, a research agent that writes reports, or an operations agent that works with your team’s tools.

Run your first agent

Create an agent from instructions, submit a task, and read its result.

Understand the building blocks

How projects, agents, releases, sessions, and runs fit together.

Connect your application

Continue conversations and steer work already in progress.

Inspect execution

Read managed run evidence, or bring traces from an agent you already run.

Who does what

Your application	Sikaru
The customer experience	The agent loop
Approval decisions	Context and retained state
Tenant and user identity	Recovery after failures
Where results are shown	Events, results, and execution evidence

How it works

Define an agent

Write its instructions and choose its tools. Each saved version is an immutable release.

Open a session

A session is one conversation. It keeps its own context and pins one release.

Send work

Each task becomes a run. The run calls only the tools the agent is allowed to use.

Read results and evidence

Read output, files, and events as the run progresses. Inspect execution evidence afterwards.

One agent serves many users. Store the session and run IDs in your application so you can reconnect, inspect results, and continue the right conversation.

What you can build

Need	Sikaru capability	Start here
Respond to a task in your product	Start a run and consume its events	Runs and tool requests
Continue work across messages	Keep a session; steer active work or queue another turn	Sessions and steering
Use an external application or API	Grant connection tools or handle product-owned tool requests	Tools and connections
Work with documents and generated files	Upload inputs and retrieve session artifacts	Files and artifacts
Work on a machine you control	Attach a native executor to a local workspace	Local compute
Run recurring tasks	Configure schedules and observe admitted work	Schedules
Assess and improve behavior	Evaluate retained traces and review candidate releases	Evaluations

Optional capabilities depend on your project’s permissions and service configuration. Start with a small task, inspect the evidence, then add tools and automation.

Choose your interface

Dashboard

Work with agents and inspect conversations.

SDKs and REST API

Integrate execution into your application in Python or TypeScript.

CLI

Terminal workflows, scripts, and local compute.

Your coding agent

Give Claude Code, Codex, or Cursor these docs and the authoring skill.

Your backend authenticates with a Sikaru project API key. You do not need model-serving credentials to run a managed agent. Pick a supported catalog model or use the project default; see Models and output.

Already running an agent elsewhere? Submit its traces without moving execution first. Trace ingestion and managed execution are separate entry points.

Next: the Quickstart, or the agent lifecycle for the complete journey.