Standards-Native Agents and Evals

Sikaru accepts behavior sources your team can review in a product repo: instructions, skills, and eval rubrics written as Markdown. These standards files give managed agents durable intent, while Experience trajectories and OpenInference-compatible traces show what happened in production. This is how Sikaru connects the evals domain to continual learning: eval rubrics define what good means, production traces reveal where behavior falls short, and reviewed improvements update the managed agent without handing Sikaru your credentials or customer-facing application surface.

File convention

A minimal agent artifact looks like this:

support-agent/
  AGENTS.md
  agent.py
  skills/
    refunds.md
    crm.md
  evals/
    refund-quality.md

Source	Meaning
`AGENTS.md`	Operating instructions for the agent.
`skills/*.md`	Playbooks, tool notes, or task-specific behavior.
`evals/*.md`	Rubrics that define what good behavior means.
`SKILL.md`	Agent Skills-style skill files for existing projects.

Load standards in Python

Python can pass Markdown paths directly from an agent artifact.

from sikaru_sdk import Sikaru, tool

sikaru = Sikaru(project="proj_123")

@tool("crm.search")
def search_crm(account_id: str):
    return {"account_id": account_id}

support = (
    sikaru.agent(
        "support",
        instructions="AGENTS.md",
        skills=["skills/refunds.md"],
        evals=["evals/refund-quality.md"],
        tools=[search_crm],
    )
    .define(outcomes=["task_success", "refund_accuracy"])
)

from_standards(".") reads AGENTS.md, Markdown files under skills/, and Markdown files under evals/.

support = sikaru.agent("support", tools=[search_crm]).from_standards(".")

Older agent.md, skills.md, and Agent Skills SKILL.md files remain supported for existing projects.

Load standards in TypeScript

TypeScript usually passes file contents from the product server or build step.

const support = sikaru
  .agent("support", {
    instructions: { path: "AGENTS.md", content: agentsMd },
    skills: [{ path: "skills/refunds.md", content: refundSkill }],
    evals: [{ path: "evals/refund-quality.md", content: refundQualityEval }],
  })
  .tools({ searchCrm })
  .define({
    instructions: "Resolve billing issues and ask for approval before large refunds.",
    outcomes: ["task_success", "refund_accuracy"],
  });

Keep traces separate from behavior

Behavior files describe how the agent should behave. Traces show what actually happened.

sikaru.traces.upload_file("./support-openinference.jsonl", dataset="prod")

This separation lets Sikaru compare intent, production evidence, and eval results without asking you to upload a runnable production harness.

Attach eval objectives

Use capture when you have static eval references and an improvement objective.

(
    sikaru.capture("support")
    .evals(
        uri="r2://acme-support/evals/support-quality.yaml",
        primary_metric="task_success",
    )
    .improve("increase task_success without increasing escalation mistakes")
    .submit()
)

What Sikaru does with standards

Sikaru reads these files as customer-authored intent, runs them through a Sikaru-managed harness, and compares future proposals against private evals and production evidence. Proposed improvements are reviewable before promotion.

Sikaru does not ask for production credentials inside Markdown files. Keep credentials in your product services and expose only tool capabilities through the SDK.

​File convention

​Load standards in Python

​Load standards in TypeScript

​Keep traces separate from behavior

​Attach eval objectives

​What Sikaru does with standards

File convention

Load standards in Python

Load standards in TypeScript

Keep traces separate from behavior

Attach eval objectives

What Sikaru does with standards