핵심 요약
요즘 AI 모델들 성능 차이가 크지 않으니 특정 모델에 종속되지 않는 에이전트 스택을 짜는 게 핵심이다. 뇌인 AI 모델을 바꿔도 내 브랜드 정보가 담긴 '하네스'만 있으면 나만의 에이전트를 유지할 수 있음. 이 하네스 기반 워크스페이스를 구축해서 클로드든 코덱스든 자유롭게 들고 다니며 쓰면 된다.
주요 포인트
- 프론티어 AI 모델 성능 차이가 좁혀져서 모델 자체보다 종속되지 않는 에이전트 스택 구축이 더 중요해졌다.
- 에이전트는 뇌(모델)와 몸(하네스), 도구, 런타임으로 구성되며, 목표를 이룰 때까지 스스로 루프를 돌며 작동하는 게 챗봇과의 차이다.
- 하네스는 브랜드 보이스나 캠페인 데이터를 갖고 있어서, 뇌(모델)를 ChatGPT 등으로 바꿔도 내 에이전트로서의 정체성을 유지해 준다.
- 로컬 에이전트(Clarko, Codex 등)는 클라우드 에이전트와 달리 내 컴퓨터의 파일과 폴더를 직접 읽을 수 있는 장점이 있다.
- 휴대용 에이전트 워크스페이스는 지식/맥락, 규칙/사양 설정, 에이전트 스킬, 전문 서브 에이전트 등 5가지 핵심 요소로 구성된다.
AI models are getting more expensive, more restricted, and honestly harder to stay committed. One week Claude feels like the best choice, and then ChatGPT gets better, and then Google shows up with something new. So, the question is no longer which model is the best because these questions matter less and less. Personally, I see the performance difference between these frontier models is getting marginal, and it's really more about how do you build your AI agent stack so you are not locked into any of them. And exactly this week a member in our community asked this same question. So, in this video, I'll show you how to build one portable AI marketing agent workspace and run it across Claude and Codex. And the same concept applies when you want to bring it to other platforms. Let's go. So, first things first, what is an AI agent really? Because understanding this concept right is really important as we design the agent stack. And here is NVIDIA CEO Jensen Huang describe it. The agent consists of model, harness, tools and skills, and a runtime. You can think of the model as the brain. The harness as the body. The tools that it uses working in a runtime. And what makes an agent instead of a chatbot is the loop. So, it takes an action, check the results, and keep going until the job is done. So, think of your AI marketing agent. The model determines its intelligence in doing marketing tasks, but the harness is what makes it your own marketing agent. It knows your brand voice, it has access to your campaign data. So, swap Claude for ChatGPT model, the marketing ability changes, but everything that make it yours, that is the harness. Now, there are two main ways AI agents can run. Cloud agents like Claude managed agent, Google Gemini Spark, Google enterprise agent platform, Perplexity computer on balance. So, these agents run on the company cloud server and don't automatically have access to your local computer. And the second way, local agents, which I believe most of us use, like Clarko, Cowork, Codex, Hermes, Open Cloud, they run on your own computer or company server. You can let them read your files and folders directly. [music] And for this video, we're covering Clarko and Codex specifically, but the general principle stay the same. So, now let's design your agent workspace. And this is the harness we just talked about. Now, of course, it can get super technical, but the simplest version to think of it is a whole project workspace with five core components. And this exact folder is what you carry to Clarko, Codex, or Antigravity, or other AI platforms. So, first, the knowledge and context, including memory. What your agent knows about you and your brand. And second, the instructions or configuration files. So, these tells how the agent should behave, rules and specs to follow in your workspace. And third, agent skills, the workflow that the agent can trigger and run. And fourth, specialized sub-agents. So, they're specialized at role definitions stored inside your harness and can be called in when needed. And lastly, tools. So, they are used in by the agents to access your real system, like your marketing analytics, CRM data. And knowing the core components is the key to making your agent system portable. Because once you see the setup as different core components, you are not asking, "How do I rebuild everything?" But really just asking, "Which are portable or not?" And we can just simply group them into three buckets. Bucket number one, completely portable. So, you can just grab them and go. Works anywhere. Bucket number two, that needs a small translation to make them work across platforms. And bucket three, isn't portable directly. So, your context and knowledge, they are bucket one since basically they are just plain files, documents, assets really. Instructions, agent skills, and sub agents, so they are bucket two because they are basically the same content, but each platform just package them a bit different. And your tools, that is bucket three since you might need to reconnect as you move between platforms. And to give you some examples, perhaps you're on a marketing team, one brand. So, you can have one shared contacts folder so the agents support from the same brand truth, and then each channel get its own project folders, paid channels, content, email, SEO, and sub agents, skills, instructions are shared across the whole team. Or maybe another scenario is you are on a functional team, say SEO. Same setup, but everything narrows to one job. And instead of channels, you split by the type of the SEO work. So, the shape is the same, but it just go deeper instead of wider. Or maybe you're a solo consultant managing multiple clients. You can organize all your clients in one workspace, each with their own contacts, but your skills and agents sit up top and shared. And if one client needs their own contacts and skills, it just live in their own folder. So, that means you inherit by default and add only