Thoughts on Agent Adoption

Introduction

Agents are changing how tech-savvy people interact with computers and the internet. Chatbots like ChatGPT, Claude, and Gemini have slowly become mainstream, but programmers and IT specialists are already moving toward agents.

Agents can do amazing things. For me, they send a daily summary of my emails, texts, and weather each morning. I use them to turn plain English into code. They're really just a different harness around the same models people use inside chatbot apps, but that harness matters a lot.

For the average person, a more powerful chatbot won't automatically make AI more useful. What will make AI more useful is personal context, tool access, and the ability to do things on its own. That's why agents feel so revolutionary. The question is: how do we make them widely adopted?

Current Agents and Drawbacks

Right now, the agent landscape is fragmented. Programming agents like Pi are better at programming than general-purpose agents like OpenClaw, which means people end up using multiple harnesses. Deep research agents are better at research than general-purpose agents too. Eventually, this has to consolidate.

The current consensus in my corner of the internet seems to be that the best general-purpose harnesses are Hermes and OpenClaw, along with variants like NanoClaw. I don't think these are the path forward. They're too challenging to set up, and even harder to set up securely.

Google's new agentic functionality inside Gemini is closer. It can check your email and modify your calendar, and it works well enough that when I need to add something to my calendar, I use Gemini voice mode to do it. I also use it for daily briefings.

But there's a problem: it's mostly good with Google services. It can't check my texts, at least on iPhone. It can't use my computer. It can't code for me in the way I want. It simply isn't there yet.

To me, the best harness right now seems to be OpenAI's Codex. With the launch of GPT-5.5, OpenAI also added computer use to it, and because it has built-in tool usage and can learn skills, it can check my email and send daily briefings easily. Of course, it can code well too. Best of all, it can do many of the things OpenClaw or Hermes can do on my computer, except it's in a sleek app that takes about 30 seconds to set up.

It's almost perfect, except for one thing.

Security Concerns

Codex is amazing, but you can't simply trust a model with access to your whole computer. There are two obvious ways to make that safer, and both make agents harder to adopt.

The first option is sandboxing. This works well, but it removes the setup convenience and gives Codex essentially the same problem as OpenClaw or Hermes. The second option is making the agent request permission for every meaningful action while you actively monitor it. That helps with safety, but it also removes much of the convenience that makes agents useful in the first place.

This is a serious problem. You could ignore it and give an agent access to everything, and for a while it would probably feel like magic. Then, eventually, it might do something dumb.

There are possible middle grounds: scoped permissions, logs, reversible actions, and trust levels for different kinds of tasks. But none of these fully solve the core issue. The more useful an agent becomes, the more power it needs. The more power it has, the more dangerous mistakes become.

This issue doesn't have an easy fix other than trusting the model, and that's hard to do. See Anthropic's alignment research and this LessWrong post on current AI alignment concerns.

Possible Widespread Adoption

Codex is already good enough to be a general-purpose agent. It might not be the best at any one thing, but it's pretty good at a lot of things. I'll continue to use oh-my-pi for coding, even though Codex is also good at coding. That's just personal preference.

As agents become more mainstream, I think many people will ignore the inherent risks because the convenience is too compelling. At that point, the main barrier to widespread adoption becomes awareness.

Codex as a brand isn't focused on a general audience. It has "code" in the name. For it to become truly successful as a mainstream agent, I think it needs to merge with ChatGPT. If that happens, it could become the first real agent ecosystem: a place where an agent can legitimately do almost anything online for you.

OpenAI could become the first agent ecosystem if it combines its apps into a superapp. MindStudio has written about this idea, and I think it's the most likely path.

Agent Ecosystems and the Future

Here are a couple of predictions.

Agent ecosystems will emerge in the same way we have tech ecosystems today: the Apple ecosystem, the Google ecosystem, and so on. Over time, those traditional ecosystems may become less important because more tasks will be done through agents.

Whether that makes existing tech ecosystems redundant depends on how open these agents become. If any agent can use Google products, iMessage, Microsoft Office, and everything else, then the old boundaries matter less. If companies gate off their software and services, then agent ecosystems will probably mirror today's platform wars.

Either way, I think agent ecosystems are the future. What matters will be what the harness can do and what the agent is integrated into.

Right now, I see two likely winners and a few possible contenders. OpenAI could win with a combined ChatGPT and Codex agent ecosystem. Google could win with a Gemini agent ecosystem. Anthropic and Apple are serious contenders too, although each lacks what the other has.

We'll see what happens. I hope it benefits us all instead of harming us all.