You haven't really seen AI yet

TL;DR

AI is more useful when it can act inside your workflow than when it talks beside it. Chat is a brain with no body.

Key points

• Chat is a brain without a body; useful for some tasks but feels clumsy.
• The model matters less than the environment it operates in.
• More capability doesn't mean more sustainability. picking the right problems matters as much as using the right tools.

I am sick of people boasting about how agents are “changing their world” and not defining what an agent is. Most people I talk to in real life are underwhelmed by AI. Meanwhile online all anyone can talk about is job replacement, industry disruption, and social crisis. I want to demystify what agents actually are and how to try them yourself. I believe what separates the disappointed practitioners and the true believers is whether you have used agents (tools that take instructions and act on your behalf inside a real workflow) for your work or not.

Once you’ve experimented with agents you realise the show is just getting started. I see agents as a digital hand helping you work. Sometimes they are not just helping, but fully doing the work. It is astonishing to see the first time you hand off a task and see it completed start to finish. To me it was much more incredible than when I first asked a question to ChatGPT in 2022 - which to that point was the most shocking experience with technology I can remember. Trying out my favourite agent the first time was as much an aha moment as an “oh s#!+” moment.

Why smart people disagree about whether AI is working

Chat interfaces with AI are just not that impressive in 2026. They offer mundane utility but they are rarely an integral part of your workflows. Chat shouldn’t be a load bearing part of your workflows unless it is a task that it is well suited for and you do it repeatably. Think writing internal blog posts, technical memo writing or brainstorming. Outside of a fairly narrow slice of tasks the Chat interface leaves a lot to be desired.

The folks that are praising AI loudly I notice are more likely to be experimenting with coding or cowork agents. These favour people in technical fields but they are definitely not limited in application here. The famous Claude Code is not just good at coding tasks for example, but it does really shine when you ask it to program for you. These tools differ from Chat because you can see them working in front of you with tangible results; files move around on your computer, documents are filled before your eyes, the AI traces its thoughts along the way so you can follow along with how it is doing. This is a step change from the question and answer rhythm in using Chat.

The people getting the most out of agents have a knack for spotting software-sized problems (tasks that are easy to solve with software because the problem is some mix of highly measurable, repeatable and connected to a digital interface). agents are perfect for solving these types of problems. Train your eye to spot these problems when you come across them in your personal (then professional) life and I promise you will start to see why everyone tech or tech-adjacent is ranting about agents.

It’s less about the model, more about where the tool sits

The model is not the be-all end-all of AI effectiveness. The model was so important in the Chat interface and it is still very important, but the tools and sandbox the model have access to are more important for seeing real productivity gains. I changed my mind on this once I started to use Codex heavily in late 2025. Codex is ChatGPT connected to your computer in a way that allows it to trigger actions inside your PC on your behalf. Codex is in other words, a harness. I see it as the harness is the body and the model is the brain.

Think about that, agency implies a subject working or completing actions on your behalf. If you have only a model (the brain) and no harness (the body); how can you expect something to be empowered to complete actions on your behalf?

A harness that can empower a model will take your context and navigate your workspace based on how the model interprets those instructions. Think: access to your database, shared document folders at work or home. The harness has permission to view these (if you allow it), the model decides how to use that information.

Chat can’t get into your work

A brain without a body has a hard time interacting with the physical world. This is how Chat feels to me lately. I am going less and less to ChatGPT (the web and mobile chat) and more to my coding agent to help solve problems I am working on. Chat is still there for brainstorming, gaming out ideas in a conversational and low friction way, but it is not my primary tool to tackle problems. Once you use an AI tool that does the work, Chat feels clumsy. You still interact with agents via chat interface (it is less pretty, but you are dictating instructions in plain language), the agent can now go do your bidding in the real/virtual world.

Think of it as a movie or play. Using Chat puts you in a dialogue but keeps you in the confines of calling out and responding to your AI. Working with agents feels much more like you are in the director seat. Agents inspect the context of the files and information you have supplied, interpret your ask and act on your behalf. This level of control over the task is where productivity gains are going to show up.

A new kind of tired

Since I’ve gotten comfortable using agents I find I am suffering from something I never would have expected. I don’t stop working. I basically finish a task and then build on the solution to make that task smoother next time or if I am feeling ambitious I tackle another project by starting from scratch. When the friction is effectively removed from completing tasks from a greenfield state (starting from scratch), you can do so much more.

This is something I am careful of, just because you have unlocked a new set of digital hands doesn’t mean it’s sustainable to constantly tackle problems to exhaustion. People who crave feeling productive all the time (hello accountants) are especially susceptible to falling into this compulsion. This has an added benefit of making me much more choosy in what I decide to focus on for automating or pursuing tasks to complete. Another thing to watch out for is thinking of your projects all the time when you are not in front of a computer screen. Not sustainable in the long run.

If you are like me you will want to throw every task at your agent immediately. Remember it’s a marathon not a sprint.

Where to start for yourself

This week I built a terminal tool for building my wedding seating plan. I don’t like killing trees and I am lousy at crafts. I asked Codex to build me an app based on a list of guests and built a whole rule set for who should sit where and a way to quickswap guests from table to table, seat to seat. So much fun, what did I do before agents?

There are several ways to do this and I think this warrants a separate writeup. My favourite harness/agent for work and home is Codex. There are three main ways to access Codex in order of difficulty for a non-technical user:

• The Codex desktop app or Claude Cowork app
• The VS Code or Cursor Codex add-on
• Embedding Codex in your terminal and interacting via command line interface (CLI)

If you’re comfortable in a terminal, both tools install in one line:

Codex with npm:

npm i -g @openai/codex

Claude Code for Mac/Linux:

curl -fsSL https://claude.ai/install.sh | bash

CLI is my favourite after getting used to it. No frills, just me and the agent. Just the way I like to focus.

If you are starting out look at the first option, if you want more power and performance out of the gate with twiddly options to tweak, try the CLI.

The future is not the Chat window

There’s going to be a lot of tension in 2026 about AI being a productivity multiple or a time waster. Point a frontier model and the right harness for your task at a software-sized problem you will change your mind.

You need to find a harness that works for your tasks. Choosing a model is not the hard part. Every big name in AI has a great model they are serving for the base subscription price. Finding a harness that leverages your business or personal context to complete tasks with as little friction as possible is going to pay much bigger productivity dividends.

The future of knowledge work is starting to feel more like managing a bench of specialized agents for your tasks. That shift is already underway for people who’ve made the jump. If you haven’t yet, the waiting room is getting crowded.

Glossary

Software-sized problem : A task that is measurable, repeatable, and connected to a digital system, making it a good candidate for automation or custom software.

Chat : Think ChatGPT, Claude, Gemini - your AI provider of choice. You interact via the chat interface. This used to feel powerful but it’s felt kind of the same since early 2025. I use it every other day, it’s great for research or to Just Google It.

agents : An interface to assign tasks to be completed on your behalf from some set of initial instructions. Agents combine a model with a harness to act autonomously. Agent is too broad a term to be useful, see below for specific examples

Different slices of agents by harness:

IDE agents : A type of coding agent. Specifically an add-on to your existing coding environment that acts as a chat window with an AI model of your choosing. The chat empowers the agent to then change your coding environment via running commands on your PC to execute actions on your behalf.

terminal agents : Alternative type of coding agent that uses a command line interface (CLI) inside your terminal. The interface is stripped down, only the essentials of the cold, efficient text interface. More fun than it sounds if you are not a visual learner.

Cowork agents : A new category, the best example is Claude Cowork. Most knowledge workers don’t perform coding tasks for a living. Cowork took the coding assistant and gave it a pretty interface for non-coders to try pointing it at their workflows. Possibly a step change for people who manipulate Excel and Word documents for a living.