Transcription
The AI industry has a big terminology problem with agents. Everything's an agent now. Chat bots, assistants, co-pilots, automations. The word has stretched so thin it means almost nothing. So, let me give you a definition that is really, really simple, but actually holds up.
An agent is an AI that can do things, not just talk. If you ask it a question and it answers, then it's a chatbot. If you assign it a task and it goes away, it executes work. Comes back with a deliverable like a spreadsheet or a document or a working application that counts as an agent. That distinction matters because it changes your relationship with the AI. If you are not having conversations and instead you're delegating outcomes, you are working with an agent.
The technical architecture behind this is simpler than the industry probably wants you to believe. Every agent consists of three components. A language model that reasons and makes decisions. Tools that let it take actions in the world, browsing websites, editing files, calling APIs, and guidance that constrains what it should and should not do. That's it. That's it. LLM plus tools plus guidance equals agent. The magic is not in any one of those pieces. It's in the combination. The sum is greater than the parts here. A language model without tools can only talk. Tools without language models require you to operate them manually. Guidance without both is just a document nobody's going to read. But if you combine all three and you get something that can receive a goal, figure out how to accomplish it and execute the steps and report back the results, now you have an agent.
I want to suggest a way of thinking about agents that will make them much, much easier to understand if you're not a technical person. I call it the little guy theory, and I think it corresponds to how a lot of us think of agents anyway, which is kind of handy. Every agent is a little guy that you hire to do a particular job. Little guy is not a genius. Little guy is not a replacement for human judgment, just a competent helper with particular skills and particular limitations. This framing matters because it sets the right expectations. You wouldn't want a new hire to have your company credit card on day one and say, "Figure it out." You'd give them a very clear assignment. You'd give them limited permissions. You check their work before trusting them with more. Agents work the same way.
The little guy framing also clarifies what you're optimizing for. You're not trying to build artificial general intelligence in your Notion workspace. You're trying to get tasks done without doing them yourself. That means reliability beats capability every single time. I would rather have an agent that correctly researches 20 companies than one that attempts to research 100 and hallucinates half the data. I'd rather have an automation that handles 80% of cases perfectly than one that tries to handle 100% and fails unpredictably so I have to manually check every single one. The goal is not to be impressed by what agents can do. The goal is not to put AI agents on your website. I, I know that's a surprise to some of you. The goal is to trust what the agent can deliver so you can delegate outcomes.
One small note on agents and pricing. If you are thinking about a hiring frame for your little guy, it helps you to understand pricing because in most cases with these agents, you're paying by the hour, the way we would pay a little guy to do work because these agents work by the token. And so it's a very similar mindset where like you're paying for the tokens that this agent will use to do the task, just as you would pay someone to help you to do a task by the hour. And so you're hiring the agent for this job. This also sets back the reliability conversation right at the forefront of your mind. Right? If you're hiring someone to do the work, you expect them to be reliable. You need to be able to expect the agent to be reliable, too, which is something that doesn't get talked about enough. I think we spend a lot more time talking about whiz-bang top 1% AI agent implementations and a lot less about very basic AI implementations that we can execute reliably that save us a ton of time and make a real difference. And that's what this video is about.
So this leads to what I call the four knobs of agent reliability. The first knob you can turn is the habitat. Where does the agent operate? Where does your little guy live? Some live on the open web, browsing websites, extracting information. Others are going to live inside your workspace, right? They're organizing and transforming content you already have. Others build software. Others connect applications and move data between them. Pick one habitat to start. Mixing them together is totally possible, but if you're just getting started, it can also create a lot more complexity than you need.
Second thing, agents need hands. What can the agent touch for tools? Read-only access is probably the safest. That means the agent essentially has a pair of glasses and eyes, and it can read stuff, but it can't write. The ability to click buttons and take actions is more powerful, but it is riskier. The ability to spend money or make irreversible changes, I would keep it off until you deeply trust the system.
The third knob that you can turn is what you would call the constraints or the guidance or even the leash for the agent. How much freedom does this agent have? A tightly leashed agent follows explicit step-by-step instructions every time. A loosely leashed agent will get goals and figure out their own approach. Beginners, I would say if you're just getting started, you want to define it as carefully as you can so that you avoid confusion and you avoid unhappy outcomes for your agents.
The fourth knob is proof. Can the agent show it did the job correctly? And so, can you specify what good looks like, what an outcome that's successful looks like that an agent needs to demonstrate? Things like providing source links or screenshots or the logs of the work or the before and after comparisons. If an agent cannot show you its work, it's really hard for you to verify its work, which means it's hard for you to trust its work.
So, with that introduction to agents, I want to give you what I would say are the best four agents that kind of fit this little guy mental model and that will help you get started if you're trying to get agents that do reliable work. And these four agents cover most of what a non-technical person needs to accomplish. I've tested a lot. They are a dime a dozen. These are the ones that you can actually use. And I'm going to tell you what they're good for.
Manis is your internet researcher. It lives in the cloud. It spins up a browser you can watch in real time. It can navigate websites the way a human would. It compiles findings into structured deliverables. Think spreadsheets or documents or slide decks. And the experience, it can be a little eerie the first time, right? You assign a task like "compare pricing and features for these top 10 competitors," and you will literally watch as it opens tabs, it scrolls through pages, it copies data into a table, it delivers a CSV 20 minutes later. You don't have to watch, but you can. And there's some proof of work there. What would have taken you 3 hours of clicking, copying, and pasting, and building a deck happens while you do other things. The free tier is going to give you, I think, 300 credits daily right now, which is enough to test it out. But paid plans get more expensive. They run from $19 to $199 a month depending on how much complexity and how much multiple little guys concurrency that you need. The key to using Manis well is specificity. Tell it what columns you want. Tell it what sources are acceptable and what format you need the output in. Vague instructions produce vague results. That's kind of a hint for most of LLM work.
Actually, I am not just recommending Manis because it's good for beginners. I am recommending Manis because it is a very powerful computer use agent for people who want to get real work done and who don't want to be in the code all the time. So there are a lot of folks who are in what I would call the professional class who use Manis as their secret weapon because Manis lets them get this comprehensive deep research stuff done in a way that they could not get any other way. And this is true even if you use something like ChatGPT deep research because you might think this is really an overlap with Gemini deep research or ChatGPT deep research. It turns out that Manis is generally speaking more complete at the kinds of deep research thinking and organization tasks and it can output in multiple formats, which is handy. So, for example, if you want to find a list of emails to reach out to about a potential fundraise and you need to reach everybody in a Y Combinator class or everybody at a particular series of funds, that is a complex task that would take a junior associate several hours. It takes Manis a few minutes and unlike ChatGPT deep research, it actually finds them all. It actually gets the whole job done and then it comes back and then it can give you a spreadsheet. It can even help to start to craft the email, etc. And so people who want work completely finished are often using Manis. And I hear back when I recommend Manis, "This is expensive." And I come back to that little agent hiring paradigm. You're hiring this agent to do reliable work just as you'd hire someone to do reliable work. If you can get in a few minutes of work all of the emails you need for a major fundraise, it's probably worth the money. And so think about the value of the work you're assigning the agent and budget accordingly.
Notion AI is another great agent. Think of it as a workspace brain. Unlike Manis, which goes out into the world to find information, Notion AI works with the content you already have. And I will just not hide the ball here. It works in Notion. So if you are not in Notion, this is not going to be as useful for you. If you are in Notion, it is tremendously useful. It works across your notes, your databases, your meeting transcripts, your project documentation. The September 2025 update introduces truly agentic work where you don't just answer questions about your workspace, but you execute multi-step tasks across your workspace. You can update a pipeline and sales estimate within Notion based on a meeting transcript automatically. For example, you can extract it to instruct it to extract every action item from your meeting notes and group them by owner and then create a task database and it will just do that. The limitation is that Notion AI comes with the business or enterprise plans because that's where they think you're going to use it. So if you're on the free plan or the plus plan on Notion, you're going to have to upgrade to get access. If your knowledge already lives in Notion, this is probably the fastest way to organize, search, and transform it. The key to using Notion AI is essentially feeding it all of your rich context. And that's why it works best with a rich existing database in Notion.
Lovable is your app builder. I've talked about it before. You describe a piece of software in plain English. "I want a personal CRM, right? I want to track my professional network with a form for adding contacts and a searchable card grid. I want to make a travel website for my family." Whatever it is, it generates a working application. It generates front-end now. It generates back-end. It generates a database. It gives you a live URL. It lets you iterate through the conversation. It helps you set up payments. This is not a toy. The applications Lovable produces use real code, usually React and Tailwind, and you can export to GitHub and continue developing yourself or even hand off to a developer later. What used to require hiring someone or learning to code yourself now requires describing what you want clearly enough to build something simple. I have been using Lovable since the beginning, and I have seen it become easier to describe what you want and get a reliable build. Paid plans will increase your message limits. Kind of like Manis, you hire what you get, right? If you want an assistant to build you a working web application, it's vastly cheaper than a developer. The key to using Lovable well is starting with a very clear mental picture of what you want and describing it precisely. The AI cannot read your mind, but it's really good at interpreting detailed instructions and Lovable keeps investing in features like visual editing that help you to more precisely realize your vision. So if you're looking to build a small application to start a business or a small application to show what's possible and demonstrate a proof of concept, Lovable is great.
The last little agent that I want to call out is Zapier. Zapier is your logistics manager. It connects applications. It automates workflows. When something happens in app A, do something in app B. When something happens in Salesforce, please put this into Slack. Right? We've had Zapier for a while. So why am I bringing it up now? Well, Zapier has added agents which add AI reasoning to these traditional workflows. So instead of rigid if-then rules, agents can analyze your incoming data, make decisions based on context, and choose appropriate actions dynamically. I would recommend starting with basic Zaps until you've built a few of them and understand how this works. If you've never used Zapier before, once you understand how they work, then start to add AI features where it makes sense. There's no point in adding an AI reasoning agent to a system that has very simple if-then rules and works better without it. The key to using Zapier well is starting with one trigger, one action, getting that working and adding the complexity of the agent when you really need it. So, for example, if you're trying to classify your incoming leads, that might take reasoning with a prompt from an agent. Maybe it works better to have an agent do that. But you might start by just seeing if you can get your leads into a spreadsheet and then you can add the classification column later with an LLM agent and see if that helps. That's an example of how I would progress through.
So theory is like this is easy to talk about, but let's try some specific examples here that you could actually do with each of these agents so you can see what I mean. Each of these does not take very long and I'm just going to give them to you briefly. They're designed for this agent and I hope they give you a sense of how easy and concrete it is to go out and get work done with agents. I don't want agents to feel inaccessible. And so this entire video is about making it easier to use them.
Try Manis. You can just open Manis and say, "Compare these top five email marketing tools for small creators in 2025. Please output a CSV with columns for tool name, starting price, free plan limits, one sentence best for description, and a source URL. Please visit the official pricing page. Please do not guess prices." And then, by the way, you can say, "I don't know what the top five tools are. Please research and determine the top five tools." Then you can just watch it work. When it delivers a spreadsheet, you can open the source links. You can verify that it got them right. Basically, it gives you a small research exercise that helps you to see how Manis works.
With Notion, find the messiest page in your Notion workspace. Whatever it is that's a brain dump in there or copied text from elsewhere. Then ask Notion AI, "Please read this page. Extract every action item into a checkbox list. Group it by person responsible. If no deadline is specified, please mark it as TBD. If no owner is clear, mark it as unassigned." This sounds really boring, but one of the most critical pieces that AI agents can help us with is our own hygiene in meetings as humans. Humans like to talk in meetings and then we don't follow up and then nothing changes. And so just little things like this, making the AI a passive, always-on feature is really helpful. And so Notion AI lets us do that. We can just define the action items, label the owners, label the due dates, and move on with our lives.
Lovable. Go to Lovable and just say, "Hey, build me up my personal CRM app. It needs a form to add a person with fields for name, company, the last time I met them, and any notes. Please display people in a card grid. Add a search bar at the top to filter by company. Please use a modern, clean design. And, uh, right now, I don't need authentication." You can add authentication, by the way, but we're just keeping it simple. Watch it build, click the preview, play around with it. You can even hit publish with it. You don't need to code. You don't need to hire someone. You just need to articulate what you want.
Lastly, with Zapier, create a new Zap. Just say, "Schedule by Zapier set to every day at 9:00 a.m. The action is send yourself a Slack message that says, 'Daily check: What's the one thing you must complete today?'" Integrate it with Slack and see if at 9:00 a.m. every day you get that little message. The most reliable workflows are just ones that are deterministic. When X happens, do Y. Now, this is not truly LLM-powered yet, but you can see how it could be, right? You can see how you could then take, "Read my last day's worth of work in Slack, turn it into a digest, turn it around, and give it to me at 9:00 a.m." Now, that's an LLM job. And so, you can easily add that complexity over the top when you're ready.
I want you to understand the core loop here. You are assigning work. You're verifying the output. You're iterating on the instructions. Everything else is refinement. So, you can start with just one agent. You can start with running a couple of missions for your agent in Manis or in Notion until you develop an intuition about what works. And then once you have something reliable, make sure you just do that use case well before you add another one. So many people try and say, "Let me do all of the things. Let me be, let me have a Claude Code instance and let me set up all of my files so Claude Code can grab them and work with them." I love that. I've done whole videos on Claude Code. It's an amazing tool. But the people who thrive with AI agents don't have to have technical backgrounds. It's just being able to articulate what "done" looks like, to be able to understand where you have unclear instructions so you can clarify them for the agent. I will be happy to do a follow-up on non-technical use for technical tools. I think that's a whole separate video, but for today, let's just focus on our team of little guys that handle work that used to eat our days. I think if we can get that far, that is already a huge win. The future is not learning to code. It's learning to delegate and having enough technical understanding of what those agents are doing using LLM and tools and guidance that you can troubleshoot. There you go. I think you have everything you need to set up your little guy and do your first agent mission. Cheers.