Managed improvement for production agents

Sikaru

You keep your product surface, tools, chat history, and approvals. Sikaru serves the harness runtime behind your API, learns from traces and evals, then proposes safer prompt, skill, memory, routing, and model releases. You approve releases before they serve traffic.

Your tools stay yoursRuntime changes stay gatedApprove the upgrade

Managed loop

Live runtime control

eval gated
12.4k runs

Observe production

Sikaru reads traces, corrections, tool results, and eval outcomes from live traffic while keeping runtime artifacts behind the API.

You keep

Feedback, eval target, approvals

Sikaru serves

Failure mining, trace clustering

eventtrace.batch.accepted
releaseprivate, versioned, approved
From code to runtime

Bring your current agent without exposing the product.

Create a project workspace

Sign up, get an organization and project, then issue the scoped key your server uses.

Capture the agent shape

Your Sikaru SDK sends source refs, tool shape, eval targets, memory hints, and compatibility constraints from code.

Stream traces continuously

The SDK streams production traces. Sikaru learns from failures and serves reviewed runtime versions behind gates.
SDK capture and trace streamingsikaru.importAgent(...) + sikaru.recordRun(...)Open platform