TL;DR:
- AI tools are shifting from "ask a question, get an answer" to "give instructions, get work done." The difference comes down to two things: agent loops and tools. The models didn't get smarter. They learned how to follow through.
- 2025 was the year of agents, even if it didn't look like it at the time. There's a 6-18 month lag between when capabilities ship and when businesses adopt them. In early 2026, Anthropic, OpenAI, Notion, Perplexity, and Meta all shipped agent tools. The convergence is real.
- You don't need to be technical. Start thinking in "projects" instead of "questions," pick a tool, and give it a shot. Those figuring this out now are building an advantage that compounds.
What Changed: From Asking Questions to Delegating Work
For decades, corporate America worked consistently across all organizations. You start a business > grow > hire employees > grow more > hire more > repeat. This was the blueprint followed by millions. It built empires, generational wealth, the Fortune 500, and the backbone of our entire economy. But a shift is happening.
AI is unlocking productivity across all industries, professions, and businesses.
In the first year, AI couldn't even browse the web. It was only as up-to-date as the most recent dataset used by the AI company. Then came real-time web browsing.
For the last two years, most people have used AI the same way. Ask a question, get an answer. Clarify, get another answer. It's a conversation. You ask, and it responds.
The promise of AI
I don't care what the AI companies tell me. I don't care how much investors pour into the AI space. I don't care how many times someone writes a post that says "[Insert AI] just killed [Insert Company]."
For years, we've been promised more efficiency and profitability. Time and again, it hasn't translated. Instead, it has resulted in millions of dollars invested in failed implementations.
They told us 2025 was "the year of agents." To some extent, it was true because we saw a lot of people use the word. But the majority of non-technical people didn't see it. Their ChatGPT, Claude, or Gemini chatbot was probably a little smarter, but largely the same. Definitely not a true agent.
Something major needed to change. And it did.
I understand it now
I believe it now because I see the shift happening. The AI companies see it too. They're putting nearly all their energy and focus into tools that run agentic models. So what's the difference between a "regular" AI model and an "agentic" AI model? It turns out, it's really just a few things that make a major difference.
(1) Agent loops: While this certainly isn't a groundbreaking idea, the deployment of AI that can run in a loop and review itself was a major development. Instead of you being the first reviewer of an AI response, AI became the first reviewer.
(2) Tools: Again, not a groundbreaking idea. If you hand a human a nail and tell them to put it into a piece of wood, we're basically useless. If you then give the human a hammer, that wood better watch out. As the creator of Claude Code has said, "The model just wants to use tools." In many cases, these tools are code and scripts.
Combining these two, you get an AI model that operates in a simple loop: gather context > take action > verify work > repeat.
Gather context: It reads the files, data, and instructions you give it.
Take action: The model determines what action it should take. In most instances, this is to call a tool. For example, search through this Word Document for [x].
Verify work: The model reviews the task from the "take action" step. If the full project plan is completed and accurate, it returns a response to you (the user). If not, it proceeds to the next step.
Repeat: If the model didn't return a response, it loops back to the "gather context" step. It will again review the context and project plan, then proceed through the rest of the loop.
We didn't need smarter models. We actually just needed to give the same models tools and allow them to run in loops.
This agentic AI model is fundamentally different from the standard chatbot (e.g., ChatGPT). Think of it this way. A chatbot is like asking a coworker a question. You ask a question and (usually) get an answer. An agent is like delegating a task. They go do it and come back when it's done. Essentially, our AI models learned how to follow through.
This is here now. In fact, it has been here for nearly a year.
The lag effect
Turns out there's a predictable period of time between the point at which something becomes publicly available and the point at which it's adopted. Occasionally, this is due to a lack of knowledge in the market. Sometimes, it's due to a technical barrier. In this case, it was probably a small amount of the former, but a lot of the latter.
Once the knowledge permeates the market and the technical barrier is removed, the adoption doesn't happen instantly. It follows an adoption curve that you've seen before. We're likely in the very beginning stages of the Early Adopters portion of the curve.
What People Are Actually Doing With These Tools
My writing is completely useless unless we understand what we should be adopting. What does using an agentic AI model look like?
Thinking in "projects" not "questions"
Google has made our brains so simple-minded. We've been subconsciously trained for decades to think in terms of a single question.
"What are the best gifts for my partner's birthday?"
Think about that question. It's only a minor subset of the ultimate "project."
When we ask for the best gifts for our partner's birthday, the project is:
- Research gift ideas that are a good fit
- Find out where I can buy each gift
- Compare prices across different stores to find the best value
- Log the gift ideas to a list with their location and price
- Review the options and buy the gift
Same with finding restaurants near me, planning trips, and more.
Stop thinking in "questions." Start thinking in "projects." The agentic AI models are best at breaking a project into tasks and completing each step along the way.
So what are some basic examples of projects it can complete?
Preparing financials from bank statements
Save your bank statements, credit card statements, invoices, receipts, and more in a folder. File format is irrelevant. Use an agentic AI model, like Claude Cowork, to read all the files and prepare a GL and income statement.
I tested this myself. Check it out here.
Preparing client deliverables with KPIs
Pull everything from last quarter out of SharePoint and prepare an executive summary with KPIs to help the C-suite evaluate business performance. Include suggestions for areas of improvement.
Updating your CRM automatically
Deploy agents that run each day by reviewing your Gmail inbox to find new contacts to add to your CRM or update information for existing contacts.
Monitoring industry news and drafting social media posts
Monitor industry sources (blogs, websites, key contacts on social media, etc.) to aggregate important industry news. When important news surfaces, automatically generate ideas for a social media post, including a first draft.
Everyone's Making the Same Bet
I'm likely just some stranger to you. Why should you believe me? Fair question. Let's move our attention to the AI industry's focus.
There are quite a few important indicators that show this is worth your time. Let's start with the first mover.
Claude Code (Anthropic)
Anthropic was the first mover in the agentic AI model space. They deployed Claude Code for public use around May 2025. It was originally shipped as a tool for software developers to write and review code. However, it's possible to run Claude Code on specific folders on your local computer to review, create, and edit files.
Claude Code is often run in the terminal, but can also be run in any IDE with the Claude Code extension. It can also be run in the Claude app.
Claude Cowork (Anthropic)
Claude Cowork was launched in January 2026. Claude Code had a bit of a moment in the last 2 months of 2025. But it's a bit scary to use for anyone who's not technical.
Claude Cowork was Anthropic's answer to that problem. It wraps Claude Code (with some limitations) into a nice interface with built-in security measures.
Since being deployed, Anthropic has been shipping features left and right. It now comes with the option to add plugins that can run tasks across all departments of a business (e.g., accounting, finance, marketing, legal). Anthropic recently shipped a feature to schedule tasks to run on a specific cadence.
Codex App (OpenAI)
The Codex app was released in February 2026. As of this writing, it's only available for macOS. Windows support should be coming soon.
In my opinion, this is a bit of a mix between Claude Code and Claude Cowork. It's a little daunting for non-technical users, but it's a much nicer interface than a terminal (again, my opinion).
It runs OpenAI's Codex AI models. As of this writing, the most powerful model is GPT-5.3-Codex.
If you're curious to read more about the Codex AI models and the Codex app, including the difference between this and Claude Code / Claude Cowork, I covered it in this newsletter.
Notion Custom Agents
Notion released a much-anticipated product called Custom Agents in February 2026. You can define a job, set a trigger or schedule for the agent, and it runs on its own across your workspace, Slack, email, and calendar. This is a great example of successfully deploying agentic AI models within an application.
I should also note here that you should be wary of what software companies call an "agent." What Notion has deployed lets agents work autonomously up to 20 minutes to complete projects end-to-end without user interaction (and often triggered autonomously).
Perplexity Computer
Perplexity Computer was released in February 2026. It just launched as of this writing. It's currently available to Perplexity Max subscribers, but will likely come to their other paid plans shortly.
This is model-agnostic, which means it deploys agents using whichever AI model is best at the specific task. It's a cloud-based system that orchestrates multiple AI models to complete projects end-to-end.
Manus Agents (Meta)
Manus launched agents in Telegram in February 2026. As of this writing, support for WhatsApp, Slack, and Discord are planned.
You send a message, and your AI agent on the other end receives it and goes and completes the request. It reports back when it's done. It functions similarly in the back-end, executing multi-step projects on a cloud-based virtual machine. You can close the app and come back later. It keeps working until it's finished.
The real differentiator here is the interface used to message your agent. Telegram is a chat app. We're already seeing a lot of this pop up, and we'll continue to see more.
OpenClaw
This is probably the best leading indicator of where the AI space is moving. OpenClaw, formerly Clawdbot and Moltbot, was released in November 2025.
It quickly caught fire because it's model-agnostic and puts the power of the agentic AI models into a text interface. Users connect their agent with tools like Slack, Telegram, and even iMessage to interact.
This was an open-source project created by a single person, Peter Steinberger. In February 2026, OpenAI hired Peter. OpenClaw remains active and open-source under a foundation.
Disclaimer: Please don't use OpenClaw unless you're extremely well-versed in AI infrastructure and IT security. It's still extremely dangerous for users who are not taking proper measures to protect their data.
The numbers back this up
I just listed 7 tools from 6 different companies/foundations. Claude Code is the only one that didn't ship within the last 3 months, and it's still constantly shipping massive updates. These companies aren't coordinating. They're following the market, and they've arrived at the same conclusion independently.
This should tell you more than any bullshit survey. It's what people want and where both individuals and businesses are finding immense value.
How to Start Using AI Agents
You don't need to know how to code. You don't need to understand automation. You just need to understand how to communicate. And you need to start thinking in "projects" instead of "questions."
Regardless of which tool you choose, there'll be a learning curve. Don't overlook the ability of AI to explain how to use itself. If you're stuck on something, ask it to help figure out how to use it. When something doesn't work how you expected, tell the AI. Oftentimes, with agentic AI models, it'll research the problem and fix it for you immediately.
I still hear so many people say they don't like AI because it gave them a bad answer in 2024 when they last used it. I can't express with enough urgency how different the models are today vs 2024, or even just 6 months ago. Especially when you're using agentic AI models in Claude Code, Claude Cowork, Codex, etc.
Your first AI agent project
Start thinking about the work you do as different "projects." This may be as complex as mapping out all your processes and preparing SOPs. Or it may be as simple as just writing down a couple things you do that you think AI could try. Remember, we're not trying to ask it questions, and we're not giving it one task at a time.
You'll have to describe everything you need the AI model to do as part of your project. You'll have to give it access to all the files and data it needs. It'll often prepare a plan for you to review first. Then you let it cook.
If it's being a dumbass, just tell it what went wrong and to try again. If you're one of those people who are scared of an AI uprising, make sure you're nice to it when you tell it how dumb it was.
Speaking from experience, you'll likely continue to tweak agent instructions over time as you develop different preferences and find different nuances. You'll probably also find more use cases for agents in your workflows as you understand them better.
Take action
We're in the early stages of the adoption curve. The lag between available and adopted is where your advantage lives. Those learning this now are working on the cutting edge. The capabilities are here, and they're only getting better. Take advantage. Take action.
Frequently Asked Questions
Q: What's the difference between an AI chatbot and an AI agent?
A chatbot answers questions. You ask, it responds, you do the work. An agent completes tasks. You give it instructions, it executes across your tools, and comes back when it's done. Think of it as the difference between asking a coworker a question and delegating a task.
Q: Do I need to know how to code to use AI agents?
No. Tools like Claude Cowork, Notion Custom Agents, and plenty of others are built for non-technical users. You describe what you want done in plain English.
Q: What are the best AI agent tools available right now?
The main options as of early 2026: Claude Code and Claude Cowork (Anthropic), the Codex app (OpenAI), Notion Custom Agents, Perplexity Computer, Manus agents, and OpenClaw (open-source, but be very careful with security). Each has different strengths depending on whether you're technical, what tools you already use, and what kind of work you want to delegate.
Q: How do AI agents actually work?
They operate in a loop: gather context (read your files, data, and instructions), take action (call a tool to do something), verify the work (check if it's correct), and repeat until the project is done. The two key developments were giving AI models tools to use and letting them run in loops instead of stopping after one response.
Q: Is it too late to start learning about AI agents?
No. We're in the early stages of the adoption curve. Most organizations are still experimenting, not deploying. The lag between available and adopted is where the advantage lives.
Q: How are accountants using AI agents?
Preparing financials from bank statements, building client deliverables with KPIs from raw data, automating CRM updates from email, and monitoring industry news to draft social media posts. These are just a few examples of multi-step projects that agents can run end-to-end with minimal oversight.
Q: What is the lag effect in AI adoption?
There's a predictable 6-18 month gap between when AI capabilities become publicly available and when businesses actually adopt them. 2025 was the year agents shipped. 2026-2027 is when firms will actually start using them. The firms that recognize this gap early and start now will build a compounding advantage.
