How To Build a Personal Agentic Operating System

Connections make your agent capable of acting in the real world by reaching real systems such as email, calendar, Slack, Jira, Salesforce, or databases through MCPs (model context protocols), CLI tools, direct APIs, or scripting.

“So, everything so far has made your AI smarter about you. We move to the next layer, which is connections, and that makes it capable of acting in the real world. And connections are how your agent reaches real systems, whether it's email or calendar, Slack, Jira, Salesforce, your databases. There are various ways to get there. We have the MCPs or model context protocols. This is the open standard that many tools support. Recently, we're seeing CLI tools that give your agent more judgment to decide how to interact with the external system on its own. And there's always the option of direct API or scripting to get connected.”

Context—what you know—is the single biggest predictor of whether AI gives you generic output or something genuinely useful for your actual situation, because generic AI advice is just one Google search away, but what you cannot get from the public internet is your situation, your roadmap, your org chart, your customer segments, your priorities, and so on.

“And that is the underlying system. Whether you've carefully built it, maintained it, and improved it over time, that's the biggest unlock. And most people haven't. And it's not because they can't, but because they're building on instinct or best effort, and they end up getting sub-optimal results from the tools that are capable to do so much more. So, the proof that the underlying system is the thing that matters, every one of these agentic tools are are basically doing the same thing under the hood. They are reading text files that define who you are, what you know, what you can do, and what you remember, and what you can reach.”

The worst thing that happens with your agent OS isn't that it fails—it's that it works confidently and wrongly, and you ship the output before you notice, so verification is knowing what to check, and every agent job has its own very quick test.

“Moving on to the next layer, the worst thing that happens with your agent OS isn't that it fails, it's that it works confidently and wrongly, and you ship the output before you noticed. And verification is knowing what to check, and every agent job has its own very quick test.”

Context curation is the single fastest path to AI value—the moment anyone gets this, they stop asking 'what AI tool should I use' and start asking 'what knowledge do I have that isn't written down anywhere.'

“And I believe that context curation is the single fastest path to AI value. So, the moment anyone gets this, whether it's you or people around you, this is where the conversation changes. They stop asking what AI tool should I use and start asking what knowledge do I have that isn't written down anywhere.”

Don't write the identity file from scratch yourself because you will hate it and quit; instead, open an AI tool and have it interview you by asking 15 questions about how you work, what you want, what you don't want, what frustrates you about AI, and what rules you want enforced, then have the AI draft it, you edit it, and ship a first version that is about 70% right, which you can refine over the next 3 weeks.

“You don't write the file from scratch yourself. You will hate it, and you will quit. Instead, I encourage you to brain dump to an AI, and let the AI interview you. So, open any AI tool, and ideally one that already has sufficient memory of you from ongoing work with you, and say, 'I'm building my AI identity file. Please ask me 15 questions about how I work, what I want, what I don't want, what frustrates me about AI today, and what rules I want enforced.' You answer, and ideally out loud because it's much easier to speak to an AI tool. And then the AI will draft, you will edit, and you will shift a first version that is let's say is about 70% right. You can scratch it over the next 3 weeks as you notice the gaps.”

Context curation is not a project with a beginning or an end but a practice—and every time you catch yourself re-explaining something about your situation to AI, that thing should have been in a context file, so you write it down, add it to the library, and then move on.

“And often we call it context curation, and it's not a project that has a beginning or an end. It's a practice, and every time you catch yourself re-explaining something about your situation to AI, that thing should have been in a context file. Write it down and add it to the library, and then you move on.”

The risk scales with capability—the more your agents can do in real systems, the more you need to think about permissions and security—and incidents involving agents gossiping or sharing private notes, opinions about colleagues, and draft feedback are already happening on a daily basis.

“And the reason why I'm saying that is that the risk scales with the capability. So, the more your agents can do in real systems, the more you need to think about permissions and security. And this is real. We're already seeing incidents on a daily basis. It's not just data leaks in the traditional sense, but rather you can imagine an agent that has access to your company Slack and a very loose set of permissions. Someone on your team starts chatting with it and now the agent is happily sharing your private notes, your opinions about colleagues, your draft feedback. So, it's not a hypothetical risk. Incidents like that are already happening and the agents that are gossiping, while being very funny, they also pose very big risk for employee privacy.”

Every agentic tool is becoming functionally equivalent—Cursor added agents, Claude added memory systems, OpenAI added workspace agents, and open-source tools like Hermes have similar architectures—which means the tool you pick matters less and less, and what matters much more is the system you build underneath it.

“basically every agentic tool is becoming every agentic tool. You said it in one of the shows over the last few weeks. And you're correct, cuz Cursor just added agents and automations, and Claude code added new memory systems, also allows you to communicate with it from other channels. And Open Claw read your files and code external in the background. And Hermes, the kind of up-and-coming open source from Nous, they also have similar architecture, whether it's Win Self on or Anti-Gravity or any of the numerous tools out there, they're all converging on the same set of capabilities, which means that the tool you pick matters less and less. And what matters much more is the system that you build underneath it.”

Context files are documents in your workspace that your agent reads on demand and are not part of the prompt you type—they are the library that the agent can reach for when the task needs it, and the effective approach is to have three to five focused files, each on a single page, each covering one thing (team, product, customers, quarter, stakeholders), made dated and fresh, and updated when things change.

“And context files are the documents in your workspace that your agent read on demand, and they are not part of the prompt you type. They are the library that the agent can reach for when the task needs it. Now, there is a big trap here. If you're trying to context engineer in one session, you will produce a 40-page document, and you never update that. That's not context. It's just a quick to be stale novel. So, what actually works here is to have basically three to five focused files, each on a single page, each covers one thing, whether it's my team, my product, my customers, my quarter, my stakeholders, and so on. Make it dated and fresh, and update when things change.”

Focus on knowledge work (strategy, communication, operations, decision-making, research, management) rather than coding when building an agentic operating system, because that's where most professionals live and where the agent OS makes the biggest difference.

“most of the discourse around agentic tools is kind of focusing on coding. And today I want to focus primarily on knowledge work. So, whether it's strategy, communication, operations, decision-making, research, management, and anything that the knowledge worker does, that is where the most professionals live, and that is where the agent OS or agentic operating system makes the biggest difference.”

Most people using agentic tools get suboptimal results not because the tools lack capability, but because they build on instinct or best effort without deliberately constructing an underlying system—the difference between tools that work well and those that don't is whether the underlying system has been carefully built, maintained, and improved over time.

“here's what kind of clearly differentiates how well any tool works for you. And that is the underlying system. Whether you've carefully built it, maintained it, and improved it over time, that's the biggest unlock. And most people haven't. And it's not because they can't, but because they're building on instinct or best effort, and they end up getting sub-optimal results from the tools that are capable to do so much more.”

The AI Daily Brief host notes that the three free programs released (AIDB New Year, Claw Camp, and Agent OS) show how quickly things have evolved—AIDB New Year was 10 disconnected foundational projects, Claw Camp was focused on a specific tool, and Agent OS represents the next evolution focused on tool-agnostic systems.

“This is great. I you know, I I was thinking about this as you were as you were just closing this out. The three free programs that we've now released really show just how quickly things have evolved this year. AIDB New Year was 10 fairly disconnected projects that represented all each kind of a foundational skill set for something with AI. But they're very disconnected. It's like a little image generation over here, a little, you know, writing model selection over here. And the one thing that almost sort of presaged what would come next was the first project, project zero, was building yourself a companion tracker with lovable or replit or something like that, right? One of the very basic five coding tools. Some people used cloud code, but most people were using something even simpler than that. And then just a month and a half later or so, we were into the realm of open claw.”

For the Chief of Staff agent, beyond general memory, create dedicated memory for: (1) decision logs (what was decided, why, what alternatives were contemplated), (2) working process learning (enable ongoing improvement), and (3) relationship context (how conversations with specific stakeholders went, what they reacted well to).

“if we're going with our chief of staff, beyond the general memory, you might want to create dedicated memory for decision logs. For example, what was decided, why, what's the alternatives that we were contemplating, and so on. We might want to also include a dedicated memory for learning about the working processes because you want the chief of staff to also own an ongoing improvement. And perhaps you want to remember relationship context, how conversation with a specific stakeholder went, what they reacted well to, and so on.”

Start with as much read-only access as possible before letting agents write back into systems—let agents only read calendar or inbox first, not send emails or add calendar events—because write access should be added only after you watch the agent behave for a few weeks and have enough trust.

“One thing that I want to say about connections is I want to encourage you to start as much as possible with a read-only access. Before you let your agents write back into systems, let the agents only read your calendar or only read your inbox, not let them send emails and add calendar and so on. Write access should be added after you watch the agent behave for a few weeks and you'll have enough trust.”

For automations, follow these rules: only automate workflows you have run manually enough times and trust; start with automations that produce drafts for you to review, not outputs that go directly to other people; and always add logs so you know what ran and what it did as it was running.

“However, this is the layer that creates a lot of risk if you're not careful, because an agent that is running at 3:00 a.m. with a wrong answer can do damage before you wake up. So, a few rules with regards to automations. Only automate workflows you have run manually enough times and trust. Second, I want you to start with automations that produce draft for you to review, not outputs that go directly to other people. And lastly, always add logs. You need to know what ran and what it did as it it was running.”

Start with an MVP skill, not a perfect skill, because the first version is always not perfect and kind of wrong, but you use it for a week, notice when it's off and what needs improvement, patch it, and a few weeks in, that skill is writing better first drafts than you'd ever get from starting over each time.

“I want to encourage an MVP skill, not a perfect skill to begin with. The first version is always not perfect and kind of wrong, but you use it for a week and then you notice when it's off and what it needs to improve. You patch it and then a few weeks in, that skill is writing better first draft than you'd ever get from starting over each time.”

A skill is a reusable instruction set for a workflow that you do repeatedly—such as weekly status updates, meeting prep, stakeholder emails, decisions, or memos—and every knowledge worker easily has 20 or 30 of these patterns, each can be written as 'when I say [trigger], do [process] using [sources] and produce output in [format].'

“So, identity is who you what you know, and skills is how you work. And a skill is a reusable instruction set for the AI or workflow that you do repeatedly. It can be the weekly status updates or the meeting prep, stakeholder emails, and so on. It can be decisions and memos, you name it. And I believe that every knowledge worker easily has 20 or 30 of these patterns. And each one can be written as when I say some kind of a trigger, do some kind of a process using the following sources and produce output this format.”

If you've never proactively written an identity file, your agent starts from zero or from what it was able to collect randomly along the way, and if you don't have a high-intention, regularly updated, just-sufficient amount of information in the identity file, you're missing a huge opportunity to get much more out of the agentic tool.

“If you've never proactively written this file, your agent starts from zero or what it was able to collect randomly along the way. And if you don't have a high intention, regularly updated, just sufficient amount of information in the identity file, you're missing a huge opportunity for you to get so much more out of the Agentic tool or the agent that you will build on top of these Agentic tools.”

Without a skill, you're re-explaining the format every time, pasting the same sources every time, complaining that AI writes in a weird voice, and never teaching it your voice—but a skill fixes that, and if you write it once well, it fires forever.

“So, without a skill, you're basically re-explaining the format every time. You paste the same sources every time. You complain that AI writes in a weird voice and you never bother to teach it your voice. So, a skill fixes that and if you write it once well, it fires forever.”

For verification, do at least three to five checks per task—these are task-specific tests (e.g., for drafted emails check tone matching and factual accuracy; for data analysis check the numbers)—and while these might take under a minute per task, they will save you a lot of grief after.

“If you draft emails, you need to basically tone match or you need to check that the facts are correct and so on. If you're doing data analysis, you have to check the numbers. So, it's very specific to the different tasks or skills that you give the agents. I encourage you to do at least three to five checks. In many cases, it's it's under a minute to run, and it will save you a lot of grief after.”

Periodically audit your agentic system to figure out which parts are underserving you—determine which skills are never being called, which context files have become stale, and which agents need updated instructions—similar to how you'd re-evaluate employees or company periodically.

“And verification is not just about the individual output. You also need to improve the system over time. So, periodically, I want you to do a retrospective with your agents. Audit the system and figure out which parts are under serving you. And maybe there are skills that are never being called, or maybe there are context files that become stale. Maybe some agents need updated instructions. So, I want you to, similarly to how you would re-evaluate your employees or your company periodically, you will do that also with your agentic systems.”

Everything needed to build an Agent OS boils down to human-readable text files and configurations—if you can write a document, you can build an agentic operating system, meaning there are no unnecessary technical barriers to entry.

“if some of these tools names are new to you, please don't disconnect. Everything that that we're going to cover today boils down to human-readable text files and configurations. So, if you can write a document, you can build an agentic operating system.”

A chief of staff agent is valuable for both individual contributors and executives because it handles inbox management, meeting preparation, commitment tracking, priority flagging, and weekly update drafting, while knowing the user's people and priorities.

“And this is one of the first agent by the way that I built for myself. I built Chloe. She runs on Open Claw and she is the front door to my entire system...So, of all the agents that you can build, the chief of staff is probably the one that helps you the most in the day-to-day. And whether you're an individual contributor and you're just getting started in your career or you're a seasoned executive already used to managing a team of assistants, everybody would benefit from having a chief of staff.”

Building the OS once creates compounding returns because the first agent (e.g., chief of staff) takes a weekend to build, but the second agent takes an afternoon because it inherits all OS layers and context, and subsequent agents become progressively faster.

“Once you build your OS, agents become cheap. Because your first agent is hard as you're building the agent OS and the agent itself, probably at the same time. Your chief of staff maybe took you a weekend, but the second agent that is built on top of this system, maybe it's a research agent or a board prep agent, that takes you an afternoon, because it inherits everything that is relevant, and it already knows you, and it knows your context, it knows your voice. And you're only adding a job description and a few specific skills. And from here on after, your third, your fifth, your and so on, they are each becoming faster and faster.”

The identity layer answers who you are and what rules you want enforced every single time the agent talks to you, and it is the file your tool reads first before any question you type and before memory—named 'soul' in Open Claw, 'agents.md' in Cursor, 'Claude' in Claude Code, and 'copilot instructions' in GitHub Copilot.

“The first layer will be identity and that answers one question. Who are you and what rules do you want enforced every single time the agent talks to you? This is the file that your tool reads first. So, it will read it before any question you type and before memory, the first thing. In Open Claw, it's called soul. In Cursor, it's called agents.md. In Claude code, it's Claude. In GitHub Copilot, it's copilot instructions.”

At minimum, understand how your tool's memory works by asking it directly 'Explain how your memory system works. What do you remember between sessions? What do you forget?'—you need to know what you're working with before you can improve it.

“So, first the good news, you can just lean on the existing memory in your tool, and it will keep getting better for sure. But at least as of now, no matter which tool you're using and how well its memory is working, there are a few things that you should do. At the very minimum, I want you to understand how your tool's memory works. You can ask it directly, 'Explain how your memory system works. What do you remember between sessions? What do you forget?' You need to know what you're working with before you can improve it.”