OKF for AI agents

What an agent can read from OKF

An AI agent needs context before it can answer, search, route, or decide whether to escalate. OKF can hold that context as Markdown files with predictable metadata.

Keep the boundary clear: OKF is a knowledge package. Tool permissions, authentication, payment actions, deletion actions, and private customer data need separate controls.

Bundle design for agent context

Minimum agent context file

---
type: Agent Context
title: Refund support routing
description: Defines when an agent can answer refund questions and when it must escalate.
resource: https://openknowledgeformat.online/use-cases/okf-for-ai-agents/
tags: [agent, support, escalation]
timestamp: 2026-06-21T00:00:00Z
---

# Purpose

Explain the stable facts, boundaries, owner notes, and related links for this concept.

# Links

- Validator: https://openknowledgeformat.online/okf-validator/
- Examples: https://openknowledgeformat.online/okf-examples/

Build a small OKF bundle for an agent

1. Pick one narrow agent task such as support routing, refund triage, onboarding, or documentation lookup.
2. Create index.md with scope, owner, safe-use boundaries, and escalation triggers.
3. Add only the workflows, policies, and references the agent needs for that task.
4. Validate one file with the OKF Validator, then validate the folder with the OKF Folder Validator.
5. Review stale context, unsupported policy claims, secrets, private tokens, customer records, and restricted operational notes before using the bundle.

Python SDK and AI agent context consumption example

AI Agent frameworks can inspect OKF YAML metadata to verify `status`, filter out deprecated concepts, and check `stale_after` timestamps before injecting concept body text into agent prompt memory (`verified`).

# ponytail: parse OKF context bundle for AI Agent prompt assembly
import yaml

def parse_okf_context(markdown_text: str) -> dict:
    parts = markdown_text.split("---", 2)
    if len(parts) < 3:
        return {"metadata": {}, "body": markdown_text.strip(), "agent_ready": False}

    metadata = yaml.safe_load(parts[1]) or {}
    body = parts[2].strip()
    is_ready = (
        metadata.get("type") is not None
        and metadata.get("status") in ["active", "verified", None]
    )
    return {
        "metadata": metadata,
        "body": body,
        "agent_ready": is_ready
    }

Why agents still need reviewed source context

1. Validate each Markdown file with the single-file OKF Validator.
2. Validate the local folder with the OKF Folder Validator.
3. Check every policy, owner, and escalation rule against the source owner.
4. Remove secrets, personal records, private tokens, and unsupported policy claims.
5. Revalidate the bundle after every source change.

Visual background

This video is included only as visual background for AI-agent context work. It is not evidence that OKF is officially required by any agent platform.

OKF is not MCP or RAG

OKF packages reviewed context. MCP connects an agent to tools or resources, while RAG retrieves relevant source material at query time. An OKF bundle can support either workflow, but it does not grant tool access, replace retrieval, or prove that an agent will use the files correctly.

• Do not store credentials, customer records, or private operational notes in OKF files.
• Do not write agent instructions that conflict with product policy or legal policy.
• Do not treat a valid OKF structure as proof that every answer from an agent will be correct.
• Do not skip escalation rules for refunds, account access, billing disputes, or safety-sensitive topics.

Related pages