Learning Library

← Back to Library

AI Agents: Action Over Conversation

Key Points

  • An AI “agent” is defined as an AI that can execute tasks and deliver concrete outcomes (e.g., spreadsheets, code) rather than merely converse like a chatbot.
  • Every agent is built from three simple parts: a language model for reasoning, a set of tools that let it act in the world, and guidance that bounds its behavior—together they enable goal‑directed execution.
  • The “little‑guy theory” frames agents as modest, hireable helpers with specific skills and limits, not as all‑knowing replacements for human judgment.
  • Properly managing agents means giving them clear, limited assignments, reviewing their work, and gradually expanding their scope, mirroring how one would onboard a new employee.
  • In practice, reliability outweighs raw capability: a trustworthy agent that consistently completes a smaller set of tasks is far more valuable than a more ambitious but error‑prone system.

Sections

Full Transcript

# AI Agents: Action Over Conversation **Source:** [https://www.youtube.com/watch?v=DAxARHKQAXs](https://www.youtube.com/watch?v=DAxARHKQAXs) **Duration:** 00:18:19 ## Summary - An AI “agent” is defined as an AI that can execute tasks and deliver concrete outcomes (e.g., spreadsheets, code) rather than merely converse like a chatbot. - Every agent is built from three simple parts: a language model for reasoning, a set of tools that let it act in the world, and guidance that bounds its behavior—together they enable goal‑directed execution. - The “little‑guy theory” frames agents as modest, hireable helpers with specific skills and limits, not as all‑knowing replacements for human judgment. - Properly managing agents means giving them clear, limited assignments, reviewing their work, and gradually expanding their scope, mirroring how one would onboard a new employee. - In practice, reliability outweighs raw capability: a trustworthy agent that consistently completes a smaller set of tasks is far more valuable than a more ambitious but error‑prone system. ## Sections - [00:00:00](https://www.youtube.com/watch?v=DAxARHKQAXs&t=0s) **Defining AI Agents: Action Over Chat** - The speaker explains that a true AI agent combines a language model, tool access, and operational guidance to execute tasks and deliver results, distinguishing it from simple conversational chatbots. - [00:03:14](https://www.youtube.com/watch?v=DAxARHKQAXs&t=194s) **Agent Pricing and Reliability Basics** - This segment explains how AI agents are billed per token like hourly work, stresses the importance of reliability, and introduces the first of four reliability “knobs” – the agent’s operating habitat. - [00:07:19](https://www.youtube.com/watch?v=DAxARHKQAXs&t=439s) **Manis AI: Pricing, Precision, Professional Use** - The speaker explains Manis’s tiered pricing, emphasizes that detailed, specific prompts are essential for effective results, and highlights its strength as a powerful research and automation tool that professionals rely on to complete complex, data‑intensive tasks far faster than a junior associate. - [00:10:53](https://www.youtube.com/watch?v=DAxARHKQAXs&t=653s) **AI-Powered No-Code App Builder** - The speaker explains how Lovable lets users describe desired software in plain English and instantly generates full‑stack, production‑ready web applications—including front‑end, back‑end, database, live URLs, and exportable code—making development far cheaper than hiring a developer. - [00:14:08](https://www.youtube.com/watch?v=DAxARHKQAXs&t=848s) **Demoing Accessible AI Agent Use** - The speaker showcases how to simplify interactions with AI agents by using Manis for automated research and Notion AI for extracting actionable items from messy notes, emphasizing ease of access and continuous assistance. - [00:17:28](https://www.youtube.com/watch?v=DAxARHKQAXs&t=1048s) **AI Agents for Non‑Tech Users** - The speaker stresses that thriving with AI agents hinges on clear instruction and delegation rather than coding expertise, and outlines how anyone can set up a simple agent to handle routine tasks while maintaining enough technical insight to troubleshoot. ## Full Transcript
0:00The AI industry has a big terminology 0:02problem with agents. Everything's an 0:04agent now. Chat bots, assistants, 0:06co-pilots, automations. The word has 0:08stretched so thin it means almost 0:10nothing. So, let me give you a 0:12definition that is really, really 0:14simple, but actually holds up. An agent 0:17is an AI that can do things, not just 0:20talk. If you ask it a question and it 0:23answers, then it's a chatbot. If you 0:25assign it a task and it goes away, it 0:27executes work. comes back with a 0:29deliverable like a spreadsheet or a 0:31document or a working application that 0:34counts as an agent. That distinction 0:36matters because it changes your 0:38relationship with the AI. If you are not 0:41having conversations and instead you're 0:43delegating outcomes, you are working 0:46with an agent. The technical 0:47architecture behind this is simpler than 0:49the industry probably wants you to 0:51believe. Every agent consists of three 0:53components. A language model that 0:55reasons and makes decisions. tools that 0:57let it take actions in the world, 0:59browsing websites, editing files, 1:02calling APIs, and guidance that 1:05constrains what it should and should not 1:07do. That's it. That's it. LLM plus tools 1:10plus guidance equals agent. The magic is 1:12not in any one of those pieces. It's in 1:15the combination. The sum is greater than 1:17the parts here. A language model without 1:19tools can only talk. Tools without 1:22language models require you to operate 1:24them manually. guidance without both is 1:27just a document nobody's going to read. 1:29But if you combine all three and you get 1:30something that can receive a goal, 1:32figure out how to accomplish it and 1:34execute the steps and report back the 1:37results, now you have an agent. I want 1:39to suggest a way of thinking about 1:41agents that will make them much much 1:45easier to understand if you're not a 1:46technical person. I call it the little 1:48guy theory and I think it corresponds to 1:50how a lot of us think of agents anyway, 1:52which is kind of handy. Every agent is a 1:54little guy that you hire to do a 1:57particular job. Little guy is not a 1:58genius. Little guy is not a replacement 2:00for human judgment, just a competent 2:03helper with particular skills and 2:05particular limitations. This framing 2:07matters because it sets the right 2:08expectations. You wouldn't want a new 2:11hire to have your company credit card on 2:13day one and say, "Figure it out." You'd 2:15give them a very clear assignment. You'd 2:17give them limited permissions. You check 2:19their work before trusting them with 2:20more. Agents work the same way. The 2:22little guy framing also clarifies what 2:24you're optimizing for. You're not trying 2:27to build artificial general intelligence 2:29in your notion workspace. You're trying 2:31to get tasks done without doing them 2:33yourself. That means reliability beats 2:36capability every single time. I would 2:38rather have an agent that correctly 2:41researches 20 companies than one that 2:44attempts to research 100 and 2:46hallucinates half the data. I'd rather 2:48have an automation that handles 80% of 2:50cases perfectly than one that tries to 2:52handle 100% and and fails unpredictably 2:55so I have to manually check every single 2:57one. The goal is not to be impressed by 2:59what agents can do. The goal is not to 3:01put AI agents on your website. I I know 3:03that's a surprise to some of you. The 3:05goal is to trust what the agent can 3:08deliver so you can delegate outcomes. 3:11One small note on agents and pricing. If 3:14you are thinking about a hiring frame 3:16for your little guy, it helps you to 3:18understand pricing because in most cases 3:20with these agents, you're paying by the 3:23hour the way we would pay a little guy 3:25to do work because these agents work by 3:28the token. And so it's a very similar 3:29mindset where like you're paying for the 3:31tokens that this agent will use to do 3:34the task just as you would pay someone 3:36to help you to do a task by the hour. 3:38And so you're hiring the agent for this 3:40job. This also sets back the reliability 3:43conversation right at the forefront of 3:44your mind. Right? If you're hiring 3:46someone to do the work, you expect them 3:48to be reliable. You need to be able to 3:50expect the agent to be reliable, too, 3:52which is something that doesn't get 3:53talked about enough. I think we spend a 3:55lot more time talking about whisbang top 3:571% AI agent implementations and a lot 4:00less about very basic AI implementations 4:03that we can execute reliably that save 4:06us a ton of time and make a real 4:08difference. And that's what this video 4:09is about. So this leads to what I call 4:11the four knobs of agent reliability. The 4:14first knob you can turn is the habitat. 4:16Where does the agent operate? Where does 4:17your little guy live? Some live on the 4:19open web, browsing websites, extracting 4:22information. Others are going to live 4:23inside your workspace, right? They're 4:25organizing and transforming content you 4:26already have. Others build software. 4:29Others connect applications and and move 4:31data between them. Pick one habitat to 4:34start. Mixing them together is totally 4:36possible, but if you're just getting 4:38started, it can also create a lot more 4:40complexity than you need. Second thing, 4:42agents need hands. What can the agent 4:44touch for tools? Readonly access is 4:47probably the safest. That means the 4:49agent essentially has a pair of glasses 4:51and eyes, and it can read stuff, but it 4:52can't write. The ability to click 4:54buttons and take actions is more 4:56powerful, but it is riskier. The ability 4:59to spend money or make irreversible 5:01changes, I would keep it off until you 5:03deeply trust the system. The third knob 5:05that you can turn is what you would call 5:08the constraints or the guidance or even 5:09even the leash for the agent. How much 5:12freedom does this agent have? A tightly 5:14leashed agent follows explicit 5:16step-by-step instructions every time. A 5:19loosely leashed agent will get goals and 5:22figure out their own approach. 5:24Beginners, I would say if you're just 5:25getting started, you want to define it 5:27as carefully as you can so that you 5:29avoid confusion and you avoid unhappy 5:31outcomes for your agents. The fourth 5:33knob is proof. Can the agent show it did 5:37the job correctly? And so, can you 5:40specify what good looks like, what an 5:42outcome that's successful looks like 5:44that an agent needs to demonstrate? 5:46Things like providing source links or 5:48screenshots or the logs of the work or 5:50the before and after comparisons. If an 5:52agent cannot show you its work, it's 5:54really hard for you to verify its work, 5:56which means it's hard for you to trust 5:57its work. So, with that introduction to 6:00agents, I want to give you what I would 6:02say are the best four agents that kind 6:05of fit this little guy mental model and 6:08that will help you get started if you're 6:10trying to get agents that do reliable 6:13work. And these four agents cover most 6:16of what a non-technical person needs to 6:18accomplish. I've tested a lot. There are 6:20a dime a dozen. These are the ones that 6:22you can actually use. And I'm going to 6:24tell you what they're good for. Manis is 6:26your internet researcher. It lives in 6:28the cloud. It spins up a browser you can 6:31watch in real time. It can navigate 6:33websites the way a human would. It 6:35compiles findings into structured 6:37deliverables. Think spreadsheets or 6:39documents or slide decks. And the 6:41experience, it can be a little eerie the 6:43first time, right? You assign a task 6:45like compare pricing and features for 6:47these top 10 competitors. and you will 6:50literally watch as it opens tabs, it 6:52scrolls through pages, it copies data 6:54into a table, it delivers a CSV 20 6:56minutes later. You don't have to watch, 6:59but you can. And there's some proof of 7:01work there. What would have taken you 3 7:03hours of clicking, copying, and pasting, 7:06and building a deck happens while you do 7:08other things. The free tier is going to 7:10give you, I think, 300 credits daily 7:12right now, which is enough to test it 7:14out. But paid plans get more expensive. 7:16They run from $19 to $199 a month 7:19depending on how much complexity and how 7:22much multiple little guys concurrency 7:24that you need. The key to using 7:26Maniswell is specificity. Tell it what 7:29columns you want. Tell it what sources 7:31are acceptable and what format you need 7:33the output in. Vague instructions 7:35produce vague results. That's kind of a 7:37hint for most of LLM work. Actually, I 7:39am not just recommending Mannis because 7:41it's good for beginners. I am 7:43recommending Manis because it is a very 7:45powerful computer use agent for people 7:49who want to get real work done and who 7:52don't want to be in the code all the 7:54time. So there are a lot of folks who 7:56are in what I would call the 7:58professional class who use Manis as 8:00their secret weapon because Manis lets 8:03them get this comprehensive deep 8:06research stuff done in a way that they 8:09could not get any other way. And this is 8:11true even if you use something like Chad 8:14GPT deep research because you might 8:16think this is really an overlap with 8:17Gemini deep research or Chad GPT deep 8:19research. It turns out that Manis is 8:22generally speaking more complete at the 8:25kinds of deep research thinking and 8:27organization tasks and it can output in 8:30multiple formats which is handy. So, for 8:32example, if you want to find a list of 8:34emails to reach out to about a potential 8:36fund raise and you need to reach 8:38everybody in a Y combinator class or 8:41everybody at a particular series of 8:43funds, that is a complex task that would 8:45take a junior associate several hours. 8:48It takes Manis a few minutes and unlike 8:50chat GPT deep research, it actually 8:52finds them all. It actually gets the 8:54whole job done and then it comes back 8:56and then it can give you a spreadsheet. 8:58It can even help to start to craft the 8:59email, etc. And so people who want work 9:02completely finished are often using 9:04Manis. And I hear back when I recommend 9:07Manis. This is expensive. And I come 9:09back to that little agent hiring 9:10paradigm. You're hiring this agent to do 9:13reliable work just as you'd hire someone 9:15to do reliable work. If you can get in a 9:17few minutes of work all of the emails 9:19you need for a major fund raise, it's 9:21probably worth the money. And so think 9:23about the value of the work you're 9:24assigning the agent and budget 9:26accordingly. Notion AI is another great 9:28agent. Think of it as a workspace brain. 9:31Unlike Manis, which goes out into the 9:33world to find information, Notion AI 9:35works with the content you already have. 9:37And I will just not hide the ball here. 9:39It works in notion. So if you are not in 9:42Notion, this is not going to be as 9:43useful for you. If you are in Notion, it 9:46is tremendously useful. It works across 9:48your notes, your databases, your meeting 9:50transcripts, your project documentation. 9:53The September 2025 update introduces 9:56truly agentic work where you don't just 9:58answer questions about your workspace, 10:00but you execute multi-step tasks across 10:03your workspace. You can update a uh 10:06pipeline and sales estimate within 10:09notion based on a meeting transcript 10:11automatically. For example, you can 10:12extract it to instruct it to extract 10:15every action item from your meeting 10:17notes and group them by owner and then 10:19create a task database and and it will 10:21just do that. The limitation is that 10:23notion comes with the business or 10:25enterprise plans because that's where 10:27they think you're going to use it. So if 10:28you're on the free plan or the plus plan 10:30on notion, you're going to have to 10:32upgrade to get access. If your knowledge 10:34already lives in Notion, this is 10:35probably the fastest way to organize 10:37search and transform it. The key to 10:39using Notion AI is essentially feeding 10:42it all of your rich context. And that's 10:44why it works best with a rich existing 10:47database in Notion. Lovable is your app 10:50builder. I've talked about it before. 10:51You describe a piece of software in 10:53plain English. I want a personal CRM, 10:55right? I want to track my professional 10:57network with a form for adding contacts 10:59and a searchable card grid. I want to 11:01make a travel website for my family. 11:03Whatever it is, it generates a working 11:05application. It generates front end now. 11:07It generates backend. It generates a 11:09database. It gives you a live URL. It 11:11lets you iterate through the 11:12conversation. It helps you set up 11:14payments. This is not a toy. The 11:16applications Lovable produces use real 11:18code, usually React and Tailwind, and 11:21you can export to GitHub and continue 11:23developing yourself or even hand off to 11:25a developer later. What used to require 11:27hiring someone or learning to code 11:30yourself now requires describing what 11:32you want clearly enough to build 11:33something simple. I have been using 11:35Lovable since the beginning, and I have 11:37seen it becomes easier to describe what 11:40you want and get a reliable build. Paid 11:42plans will increase your message limits. 11:44Kind of like Manis, you hire what you 11:46get, right? If you want an assistant to 11:48build you a working web application, 11:50it's vastly cheaper than a developer. 11:51The key to using lovable well is 11:54starting with a very clear mental 11:55picture of what you want and describing 11:57it precisely. The AI cannot read your 11:59mind, but it's really good at 12:01interpreting detailed instructions and 12:03lovable keeps investing in features like 12:06visual editing that help you to more 12:08precisely realize your vision. So if 12:11you're looking to build a small 12:12application to start a business or a 12:14small application to show what's 12:15possible and demonstrate uh a proof of 12:18concept, Lovable is great. The last 12:20little agent that I want to call out is 12:22Zapier. Zapier is your logistics 12:24manager. It connects applications. It 12:26automates workflows. When something 12:28happens in app A, do something in app B. 12:31When something happens in Salesforce, 12:32please put this into Slack. Right? We've 12:35had Zapier for a while. So why am I 12:37bringing it up now? Well, Zapier has 12:40added agents which add AI reasoning to 12:42these traditional workflows. So instead 12:44of rigid if then rules, agents can 12:46analyze your incoming data, make 12:48decisions based on context, and choose 12:51appropriate actions dynamically. I would 12:53recommend starting with basic Zaps until 12:56you've built a few of them and 12:58understand how this works. If you've 12:59never used Zapier before, once you 13:01understand how they work, then start to 13:03add AI features where it makes sense. 13:06There's no point in adding an AI 13:08reasoning agent to a system that has 13:11very simple if then rules and works 13:13better without it. The key to using 13:14Zapier well is starting with one 13:17trigger, one action, getting that 13:18working and adding the complexity of the 13:21agent when you really need it. So, for 13:23example, if you're trying to classify 13:25your incoming leads, that might take 13:27reasoning with a prompt from an agent. 13:29Maybe it works better to have an agent 13:31do that. But you might start by just 13:34seeing if you can get your leads into a 13:35spreadsheet and then you can add the 13:37classification column later with an LLM 13:39agent and see if that helps. That's an 13:41example of how I would progress through. 13:43So theory is like this is easy to talk 13:45about, but but let's try some specific 13:48examples here that you could actually do 13:51with each of these agents so you can see 13:53what I mean. Each of these does not take 13:55very long and I'm just going to give 13:56them to you briefly. They're designed 13:58for this agent and I hope they give you 14:00a sense of how easy and concrete it is 14:03to go out and get work done with agents. 14:06I don't want agents to feel 14:08inaccessible. And so this entire video 14:10is about making it easier to use them. 14:12Try Manis. You can just open Manis and 14:15say, "Compare these top five email 14:17marketing tools for small creators in 14:192025. Please output a CSV with columns 14:22for tool name, starting price, free plan 14:24limits, one sentence best for 14:26description, and a source URL. Please 14:28visit the official pricing page. Please 14:30do not guess prices. And then, by the 14:32way, you can say, I don't know what the 14:34top five tools are. Please research and 14:36determine the top five tools. Then you 14:38can just watch it work. When it delivers 14:40a spreadsheet, you can open the source 14:42links. You can verify that it got them 14:43right. Basically, it gives you a small 14:45research exercise that helps you to see 14:48how Manis works. With Notion, find the 14:50messiest page in your notion workspace. 14:52Whatever it is that's a brain dump in 14:54there or copied text from elsewhere. 14:56Then ask Notion AI, "Please read this 14:58page. Extract every action item into a 15:01checkbox list. Group it by person 15:03responsible. If no deadline is 15:05specified, please mark it as TBD. If no 15:08owner is clear, mark it as unassigned." 15:10This sounds really boring, but one of 15:12the most critical pieces that AI agents 15:15can help us with is our own hygiene in 15:18meetings as humans. Humans like to talk 15:20in meetings and then we don't follow up 15:22and then nothing changes. And so just 15:24little things like this, making the AI a 15:27passive always on feature is really 15:30helpful. And so Notion AI lets us do 15:32that. We can just define the action 15:34items, label the owners, label the due 15:36dates, and move on with our lives. 15:38Lovable. Go to Lovable and just say, 15:40"Hey, build me up my personal CRM app." 15:42It needs a form to add a person with 15:44fields for name, company, the last time 15:46I met them, and any notes. Please 15:48display people in a card grid. Add a 15:50search bar at the top to filter by 15:51company. Please use a modern clean 15:53design. And uh right now, I don't need 15:55authentication. You can add 15:56authentication, by the way, but we're 15:57just keeping it simple. Watch it build, 15:59click the preview, play around with it. 16:01You can even hit publish with it. You 16:03don't need to code. You don't need to 16:05hire someone. You just need to 16:06articulate what you want. Lastly, with 16:08Zapier, create a new Zap. Just say 16:10schedule by Zapier set to every day at 16:139:00 a.m. The action is send yourself a 16:15Slack message that says daily check 16:17what's the one thing you must complete 16:18today. Integrate it with Slack and see 16:21if at 9:00 a.m. every day you get that 16:23little message. The most reliable 16:25workflows are just ones that are 16:27deterministic. When X happens, do Y. 16:30Now, this is not truly LLM powered yet, 16:33but you can see how it could be, right? 16:35You can see how you could then take read 16:38my last days worth of work in Slack, 16:40turn it into a digest, turn it around, 16:42and give it to me at 9:00 a.m. Now, 16:44that's an LLM job. And so, you can 16:46easily add that complexity over the top 16:48when you're ready. I want you to 16:49understand the core loop here. You are 16:51assigning work. You're verifying the 16:53output. You're iterating on the 16:55instructions. Everything else is 16:57refinement. So, you can start with just 16:59one agent. You can start with running a 17:01couple of missions for your agent in 17:03Manis or in notion until you develop an 17:05intuition about what works. And then 17:07once you have something reliable, make 17:09sure you just do that use case well 17:12before you add another one. So many 17:13people try and say, "Let me do all of 17:15the things. Let me be let me have a 17:18Claude code instance and let me set up 17:21all of my files so Claude Code can grab 17:23them and work with them." I love that. 17:26I've done whole videos on cloud code. 17:28It's an amazing tool. But the people who 17:30thrive with AI agents don't have to have 17:32technical backgrounds. It's just being 17:34able to articulate what done looks like 17:36to be able to understand where you have 17:38unclear instructions so you can clarify 17:41them for the agent. I will be happy to 17:43do a follow-up on non-technical use for 17:46technical tools. I think that's a whole 17:48separate video, but for today, let's 17:50just focus on our team of little guys 17:52that handle work that used to eat our 17:54days. I think if we can get that far, 17:56that is already a huge win. The future 17:59is not learning to code. It's learning 18:02to delegate and having enough technical 18:05understanding of what those agents are 18:07doing using LLM and tools and guidance 18:10that you can troubleshoot. There you go. 18:12I think you have everything you need to 18:13set up your little guy and do your first 18:16agent mission. Cheers.