Ep 871: Desktop Agent Lingo Simplified: Goals, Loops, Plans, Subagents and how it works in Codex and Claude Code (Start Here Series Vol 30)

Welcome to the Everyday AI Podcast. My name is Jordan Wilson, and for the past 3 and a half years, we've put out more than 800 episodes. Yet one of the most common questions I get, I didn't really have an answer for. Where do I start on the Everyday AI Podcast? And that's why we started the Start Here series. And with fall now back in full swing, the Everyday AI Podcast is going back to school and playing back the entire Start Here series. From front to back. We've hit pause on our normal Monday to Friday programming to run back our most popular series ever for the next 30 days. We made the Start Here series for beginners and AI champions alike. So whether you're just trying to get a grasp on large language models or grappling with the best coding harness for multi-agentic workflows, the Start Here series covers it all. Plain language, no jargon, and easy to follow along each day. So make sure to subscribe to the podcast and check back each day for new insights. Day by day, the series is a culmination of spending more than 10,000 hours covering generative AI over the past 3 and a half years. So you don't want to miss a single episode of the Start Here series. Let's get into it.

Remember back in 2024 when knowing how to use a custom GPT was a differentiator? Or in 2025 when knowing the difference between a skill and a project in Claude could be a competitive advantage for your company's AI efforts. Well, those days are gone, and so is most of the useful lingo, because controlling front-end chatbots and models and modes are now table stakes. The differentiator now is controlling long-running desktop agents. And it sounds easy in theory until you realize that the language you've been building up on the front-end AI chatbot era and the language of the long-running autonomous agents are not exactly compatible. So today we're giving you the primer and establishing some baseline vocabulary and concepts you'll need heading into 2027 when long-running AI agents become the new norm. So here is the big picture. Agent vocabulary is now an essential skill set. And yeah, it's changing all the time, which is one of the reasons why we do this thing every day. But agents can obviously read and write files. They can run tools, call their own, create their own apps and plugins. They can fix mistakes and work for hours unattended, which is both a good and a bad thing depending on how active you are in your agents. And every new term now names a problem you only hit once an agent run— runs long, right? So, so, so much of the previous terminology that we use with AI chatbots, right? You had an instant feedback loop for yourself whether it worked or not. And that is kind of gone. You really have to be paying attention. And not knowing the words now, well, means you might set a vague goal, you might put up weak guardrails, and that run that you think might fix a problem as you go take a walk might not really go anywhere. And learning this new agentic language is becoming as essential as learning how to prompt, as that was helpful. So on today's show, we're going to learn what loops, goals, plans, and subagents actually mean without the jargon. In. Uh, we're gonna go over how all of these terms work in Codex and Claude Desktop so you can know how these features work together. Uh, you're gonna learn why fluency in these terms is the skill separating operators from spectators, and you're gonna know the mental model that makes the whole vocabulary click starting now. All right, if you're new here, Welcome. This is the Start Here series. The Start Here series is part of Everyday AI's ongoing effort to give, whether you're a new listener or a seasoned AI expert, to give everyone an essential podcast series to both learn the AI basics and double down on your AI knowledge. So if that's what you're trying to do, sweet, me too. Let's do it together. Make sure to go to starthearseries.com. That's going to give you exclusive access to our inner circle community. And you're going to be redirected straight into our Start Here series space, which has a playlist for all of these episodes, all in order. So it's easy to go through them all as well as you can read about them all on the page and connect with other people who are along the journey with you. So if you missed our last Start Here series, that was volume 29, where we talked about the open source surge and if models like GLM 5.2 Uh, make open source and enterprise priority. But today we are talking about the desktop agent lingo simplified. So first, let's zoom out completely.

All right.

So if you are not someone that's using Codex or Claude Desktop or Anti-Gravity or Cursor, uh, some of this might not make a ton of sense and that's okay. And maybe this show is more for you than anyone else, because if you are using something like Codex or Claude Desktop every single day for hours, this episode will probably be a review at best, but I think still helpful. So if you're like, okay, I don't use these tools, no, you need to listen up and you should start using them. But I want to talk about the shift, obviously, from the AI chatbot that's just very reactive versus the proactive autonomous desktop worker. And really what separates them is the harness, right? So sometimes when you talk about a model, you talk a lot about the harness or where it lives and how it accesses, how they access all of these tools, right? So let's just use an example, Codex from OpenAI, right? It's my favorite harness. It's the one I use most, but a lot of people don't even know you can use other models inside Codex. Codex is the harness, right? You can't do that with Claude Code. You can do that with Antigravity, Codex, and some others, Cursor as well. So the harness is how all of these agentic tools come together and they can work over long term, right? When you see all these stories, you're like, oh, this person had an agent running over the weekend or overnight. Well, that's really made possible by the harnesses that these companies create that give the models and the tools essentially a sandbox to do all this. And, you know, it's kind of, uh, today's episode is really a primer on understanding what happens under this, the hood of the harness, so to speak. So we're not going to get super technical. This is, uh, for beginners and maybe intermediate people as well, but that's what we're truly trying to understand. And like I said, even if you're not using these desktop tools now, I think most people will be using them, uh, come 2027, right? We've heard from Microsoft that they're coming out with their super app Right. We tackled super apps on the show and on the Start Here series. What episode was that? Bringing that up here. That was our episode 799 or volume 28 of the Start Here series. So, you know, Microsoft is bringing out a super app. So these super apps or these agentic harnesses, right, if you want to get a more technical term, essentially they are a much more powerful version of a web-based chatbot that can run autonomously on your computer. You can run them in loops. You can run goals. We're gonna go over all these things. You can schedule automations. So the big difference is, well, it can act autonomously. It can share the context across different chats. And it can read and write to your computer. So any file, just like a human really would, it can use your actual computer with computer use. It can Use your browsers, right? Has built-in browsers. So, you know, definitely go listen to the AI Super Apps episode if you want a little bit more primer. But let's get straight into it now. So let's talk about plans. And I think plans are actually one of the more underrated features of these harnesses, specifically in Claude Code and in Codex, right? Every once in a while, I'll ask people how they're using these different harnesses and I— people don't really talk about or use plans as much as they should. So a plan essentially just reveals the route before an agent asks, uh, or acts. So a plan shows the intended steps before, you know, the files, the tools, the app changes, all of those things. And planning exposes the assumptions, uh, right? So likely files, approval points, and verification steps. So plan modes. matter because desktop agents can change work fast and they can work for a very long time, right? So it's kind of like a blueprint for a building, right? You wouldn't just— if you had unlimited resources and you wanted to build a building, you wouldn't just go to someone and be like, hey, go build a building. Or if you're building a custom house, you wouldn't just be like, hey, I want a house and make it awesome, right? You'd probably sit down with an architect and you'd go over the blueprints, the floor plans, zoning restrictions, requirements, all those things, right? It might sound tedious, but you're probably going to get a much better result in the long run if you sit down and have the conversation with said architect or general contractor, right? Whatever it is, that's the same thing a plan is. It is literally a plan. So a lot of people will usually just point, you know, Claude Desktop or Codex to a folder or share a little bit of context and say, get to work, buddy. Not a good idea. So the plans are essential. So Codex, right, uses the kind of plan, pair, and execute as collaboration gears. So it reads, it analyzes, and then it proposes things to you, and then it waits for you to approve it before implementation. So similarly, that's how Claude Desktop works. It actually works out nicely that these work very similarly in Claude Desktop and in Codex. So if you're actually using these and if you want to see like, how does this work? Where's the button? Right? Codex has a plan button. I don't believe Claude Desktop has it, but you can invoke them each the same way, which is just a backslash and then plan. Right? And then they'll kind of walk you through, and then you will approve the plan. So in the same way how I am very adamant telling people, like, make sure you read the chain of thought after you've gone through, you know, and gotten an output out of a large language model. I am that much in favor of using plan mode. It is the equivalent on the front end, right? So if you've never used these harnesses, think of something like Google's— Google Gemini's deep research actually does a really good job of this, right? Before you go out and do a deep research with Google Gemini, it actually gives you a plan to approve so you can see it. And if something's wrong, you can modify it. And this is really important because when we talk about the shift from token maxing to token efficiency, Right? Depending on how you set these agents up and what kind of task you're giving them, they might run for a long time. And, you know, if you are using these on a company plan, chances are you're paying via the API. So not having a plan, in the same way telling a builder to build you an amazing house, can be a very expensive step to overlook. Plan mode, you have to think it's not slowing you down. A lot of people are just like, I want to build, I want to break stuff, I want to burn tokens. No. It's a guardrail, right? Not a slowdown. So planning separates thinking from doing, but it's not, you know, it's not enough because risky work still needs, right, that read-only access, the work trees, the checkpoints. And then once your route is approved, you know, then the permissions kind of define the real blast radius. So don't just rush into giving a long-running agent keys to the castle or, you know, giving it access to just run for hours and burn through your API bill. All right, next let's talk about goals. All right, and I, I kind of put these in an order that I think might make sense for most. Um, I generally will start with a plan, and the way I actually do it is I use a plan and then, uh, parlay, uh, the findings of that into a goal. So it's a little confusing, and, uh, Claude Code and Codex work a little bit differently. Uh, uh, uh, in for plan mode, it will still work through your plan, but a lot of times it will stop once it's gone through the steps. Uh, so the biggest difference between a plan and a goal is on the front end, uh, a plan literally just outlines it and you can see, uh, Codex or Claude Code work through each step and there is a visual indicator, which is great. Goal is a little bit different. Goal is you literally give it an end goal. And it will not stop and, and sometimes loop until it hits that goal. Right? So, a lot of times you have to be careful using goal if you don't go through a plan mode first, or if you don't go through an old school best practice prompt engineering, context engineering like we used to teach with Prime Prompt Polish to really share that context and have an understanding with the model of what ultimately the output looks like. So if you just blindly, and a lot of people do this, right? They just blindly throw a bunch of context. at Codex or Claude Desktop, and they give it a goal and it will keep going until it hits that goal. And if that goal is maybe unattainable, yeah, that's where you can have it work for hours on end or maybe a day or longer. And all of a sudden, yeah, your bill's through the roof. So let's talk a little bit about goals. So it's really defining the finish line before you start any motion. And like I said, I always go through a planning phase first. I usually make the plan kind of vague, uh, to work through it. Uh, let, you know, Codex or, uh, Claude go through and do some work. Then I'll see, you know, there's always gonna be shortcomings the first time you give something, uh, something to someone. So I'll see where it went wrong. I'll go through, read the summarized chain of thought, then I will kind of reprompt it and then make that as a goal. All right. And sometimes I'll literally just copy and paste the original plan, make some modifications, obviously, because some things on the list or on the steps of the goals are gonna get crossed off. Some things aren't. So I usually do a lot of manual nitpicking transitioning from a plan to the goal. But the goal is the outcome an agent checks across the long run. And strong goals specify the audience, deliverable, source material, and the done condition, which is why I think it is usually best in general terms for, you know, general knowledge work. And again, I'm coming at this from like a general knowledge work. Yes, I do some coding software dev stuff, building myself You know, cool projects and products, extensions, all those things. But I am talking about this through general knowledge work. So even if you're not, you know, writing software or, you know, vibe coding something, let's just say you're vibe working. I still think that this holds true. So weak goals make plans, loops, and subagents drift quietly, which I think it's important to go through, especially if you're talking about a project that you want to do, right? If you're just trying to knock off a task that might normally take you or a single agent on the web, you know, 5, 10 minutes to go through, I don't think you necessarily have to go through the plan and the goal in combination. But if you're talking about knocking out an actual project, which is what these systems are capable of doing right now, that might take a human 2, 3, 4, 5 hours or 2, 3, 4, 5 days, that's when I think it, you know, when you're talking about your ROI and what you have to invest the front end. This is one of those things where you're going to have to build the bridge in order to save the 4 days driving around the mountain. Uh, so goals differ a little bit across Codex and Claude. So Codex goals are a little bit more persistent and editable and also visual in the composer. There's a nice little toggle, uh, button in the goal-first threads, uh, in Codex, kind of keep the objective across turns and sessions. Uh, Claude code, uh, via the desktop version, not the CLI. It does still support the backslash goal in the same way, but Cowork is a little bit differently— works a little bit different. That's the other thing people, you know, and for me, I personally use, you know, Codex, and I do have to tell people this because if not, it would be extremely irresponsible of me. You know, the big difference between Codex and Claude Desktop is Claude Desktop is fragmented. There is a chat, a cowork, and a code tab, and those tabs have no clue what the other is doing. So there's, you know, certain things that work really well in cowork Goal, such as the visual indicators that don't always work as well in the code version of Goal. So, you know, even as we go over individual features, because Claude Code by default— or sorry, Claude Desktop by default is siloed between the chat, the cowork, in the code, you know, even things like goal work a little bit differently in Cowork and in Claude Code. All right. Another thing to keep in mind, a goal is not just a step-by-step prompt, right? So over-specified steps can actually fight the agent's own planning process. So you don't have to know everything going into it, right? You don't have to like be like, oh my gosh, I can't use this plan thing and the goal thing because I don't have experience. I'm not a software engineer. Absolutely not. Right? You have to be able to communicate and understand what you want. And then work with an agent to, you know, go through the steps to get there. But underspecified outcomes make the agent invent missing, like, success criteria, and that's what you absolutely want to avoid. All right. So now let's talk about loops, right? So loops are kind of the newish trend, right? Even though we had the Ralph loops, you know, many, many quarters ago, now loops are making their viral reappearance again. So loops essentially can turn plans, approved plans, into agent work, right? So similarly, like a heartbeat if you're using OpenClaw, right? So think of it like this. It's something that you can schedule to happen over and over and over and over again. So a loop means, you know, the agent is going to observe, plan, act, check, adjust, and repeat. It is literally A loop, so a chatbot would run once while a long-running agent can loop many times if you give it that instruction. So the loop only becomes useful when the agent verifies each step. Otherwise, you're just burning tokens like you're still trying to climb the internal meta, you know, token burning leaderboard. So, you know, loop review looks a little bit different in each product. So Codex runs every task loop inside a project organized throughout. So, you know, you can also build loops as skills or automations, right? So you can just tell them, hey, here's what I want you to do, right? Every, you know, every hour I want you to go through, you know, triage my email, my calendar, and my Drive, you know, respond to any emails that I might want to update any decks. And I want you to do this every hour. I want you to do this, you know, every day. Right. So that's an example of a very oversimplified example of what could be considered a loop. Right. So Claude Cowork, again, a little bit different. It shows the steps with citations to the files and the messages, which I really like. I like that, that piece in Cowork a little bit more than I like it in Claude Code because it has that dedicated side panel on the right. So even if you are running technically a loop, you can see those steps visually get checked off. Claude Code surfaces the plans, the tasks, the subagents, the diffs. In the progress panels. The biggest thing to talk about when, you know, talking about loops is having very clear verification, right? Whether you want to run something a first time, right? I would literally— and I've done this before— you can work through, just go through a natural language, work with Codex, work with Claude Code, and say, hey, I want to create a loop. I want to ultimately save this as a skill or an automation, or, you know, usually both. Save it as a skill and an automation that you can schedule and run. But say, hey, instead of giving you the full loop, I want to work with you through this. Here's ultimately what I want to do. I want to check this website twice a day. This is my whole company. We live and die by this website, right? It could be industry news, could be, I don't know, stocks, finances, et cetera, right? But think if you are someone that has to pay very close attention to the market, something in healthcare, I don't know, FDA regulations, and there's all these things pinging all the time, right? Maybe you want to create a loop that you train or you help build a skill set with Codex or Claude Code that goes in there every so often. You know, you see what's new, you run it through your context, your decision-making process, et cetera. Work through those steps in the loop one by one. So the, so the harness or the model in the harness can understand what verification looks like. What success looks like at each step of the loop. If not, bad loops, bad, right? Bad loops turn into producing, uh, you know, overly polished work that just is maybe wrong. Because when one loop gets overloaded, then sub-agents are maybe going to split the work. All right, so loops are great. You can save them as a skill in automation. Think of it similar to like a heartbeat. Uh, if you've used OpenCLAW, the great thing, all it takes is natural language. Uh, right? I do think in this aspect Codex is a little bit better at setting these up. I think that they're, uh, because not being fragmented, uh, it's a little bit easier to just chat, uh, with Codex in natural language to set up those loops. All right, and then last but not least, we have sub-agents. So, all right, sub-agents like loops are where things get very interesting and potentially dangerous if you not being hands-on. So if you're being laissez-faire, human in the loop, and setting all these things off, I wouldn't recommend going too heavy in the paint on loops and subagents. It will get expensive quickly. If you are very hands-on, I think loops and subagents are great. So here's what subagents are. Well, they're just helper agents with focused assignments and separate context windows. So they help with parallel work, not just vague requests to think harder. The real value, though, is context hygiene before the consolidation begins. So subagents control kind of a different aspect of a job. So one easy thing, right? An example of how I use subagents in Codex and in Claude. So you can just use them. By invoking them. Say, use subagents for this. A lot of people are like, how do you use subagents? Well, you natural language, use subagents for this. So as an example, one thing I like to do is I like to build with Codex. All right. So I build with Codex. A lot of people are like, oh, Claude is better at front end. Yeah, it's way better at front end. But I don't know, Codex has this thing called the most powerful AI image model in the world. So start with the AI image gen in Codex. Use the front-end skill, takes 3 seconds, and you generally will have a much better front end if you are building, you know, a piece of software. That's just what I'm using as an example, right? If you use the image gen and the front-end design skill, you're gonna have way better than what you would get out of Claude Code if you do that. If not, obviously Claude Code is gonna be better. But one reason, or one way I use subagents, I, I love Claude's ability to really investigate and use subagents across multiple things. So I'll sometimes, I'll assign You know, I'll say, hey Claude, I want you to do these subagents. I want you to look at, you know, one, look at the entire repo. Number, you know, 2, look at the feature set. 3, look at the language, right? So then each subagent can really have a more defined and refined kind of task, right? So it's the same thing. If you just walked into a room of 20 employees and they're there to help you and you say, hey, 20 employees, Go help me with this project. Here's where we're at. Have fun. Right? They might decide upon themselves, okay, let's split this up. But if you say, hey, designer, go look at the design. Hey, copywriter, go look at the copy. Hey, engineer, go look at the, I don't know, the security on the backend, make sure everything's, you know, tightened up. So all you have to do is, well, say, go use subagents. Obviously, these models are smart enough where you can just say that broadly, and they will, you know, usually, uh, you know, assign one sub-agent to a specific task. Uh, usually I'll just broadly say, go use sub-agents and poke holes in ABC, and then I'll see exactly how they did it. Uh, Claude and Codex allow you to click, uh, onto a sub-agent and see exactly what they're working on in real time, which is cool. I love how Codex names them. Uh, the Claude, uh, names— I forgot how they named them. But, um, you know, usually I'll do a quick general sub-agent run if it is a big project that's super important, right, that I can spend kind of the token budget on. I'll see what they did, what went wrong, what didn't, uh, you know, how it can be improved. And then I'll probably assign roles the second time I have another group of sub-agents do it. Sometimes I'll have sub-agents work on the front end before the work actually starts to scope it and make sure it makes sense on where they're spending their time. Sometimes I'll have them work on the back end after the is done, or a combination of the both. And the best thing is, well, you can just tell Codex or Claude, like, hey, have subagents work on the front end before you start this project to make sure everything is correct. Then go do the work. Then have a separate group of subagents really tear it apart on the back end.

Right?

That's all you really need to say, and they will do that. It, it is kind of like this, having this parallel work stream that can disagree, duplicate, you know, miss shared constraints, and that's great. And then the main agent can compare findings, resolve conflicts, synthesize, I've even set up a skill that has like a 3-tier kind of agentic management system. So I don't even have to go through and type this, right? So if I know a project is pretty big, or if I'm going to be working, you know, it's like, hey, I'm setting this off to work overnight. It's an ongoing project and I'm not going to go through the planning and goal. I know I have this, you know, sub-agent system where I, you know, say, hey, this one's going to go in and be a contrary or contrarian and, you know, tear apart all this and second-guess every single thing that we do before it goes to production, whatever it may be. Uh, so that's just, uh, a way to go through subagents. All right. So now as we wrap up, the real constraint here is actually your machine. This is the big thing to keep in mind because your machine has to be on. It has to be, uh, right? You have to have these programs open. So yeah, you can do all these great things like control your computer, right? And, and control the browser and think of that as well as we talk about So, you know, the desktop lingo is it's literally as if you were talking to a human sitting in front of the computer, but it has to be on. Right? A sleeping laptop can break remote steering. You know, if you run out of power, if your internet goes off, everything, well, from an agentic perspective, stops. Where that's a little bit different, the advantage of using on the front end chatbot, those, you know, singular tasks, those singular one-off prompts can continue to go on. But the beginner mistake is not managing anything. Right? The beginner mistake is being like, wow, these things are super powerful. These agents, I can just go in and, you know, drop it some context and say, go do this work. And wow, look at it. You're gonna get just slop work almost every time. Right? Because also bad handoffs say, research this, organize that, make that better. And then you just use the output. Right? Good handoffs are, well, defining the goal, the plan, the permissions, the workflow, the verification, the subagents, the loops. Right? That's the differentiator now, not just saying, oh, I use Codex. Oh, I use Claude Code Desktop. Right? The handoff turns the agent lingo into repeatable desktop delegation. So like I said, the new skill is number one, understanding how to talk to your agents, but then using those skills to supervise work, not just prompt. Right? The agentic layer is growing thicker and thicker by the day. That doesn't mean the human layer Both, you know, I call it the, you know, the agentic human sandwich, right? We are the bun. We are providing the input and the context on the front end, and then ultimately the verification. But that doesn't mean we're hands-off in the agentic layer. We have to constantly be monitoring, improving. So we don't just, you know, blindly let these agents loop and, you know, blindly assign work to sub-agents, right? We can just spend more time as the bun and far less time as the meat, to use my old saying there. All right, I hope this one was helpful going over the basics of desktop agent lingo, simplifying goals, loops, plans, uh, and subagents, and a little bit of how they work in, in Codex and in Claude Code on the desktop. So if this was helpful, please go to starthearseries.com. If you haven't already, make sure to subscribe to the podcast. I would really appreciate that. Thanks for tuning in. Hope to see you back tomorrow and every day for more Everyday AI. Thanks, y'all. Bye.

And that's a wrap for today's edition of Everyday AI! Thanks for joining us! If you enjoyed this episode, please subscribe and leave us a rating. It helps keep us going. For a little more AI magic, visit youreverydayai.com and sign up to our daily newsletter so you don't get left behind. Go break some barriers, and we'll see you next time! Bye!