There is only one agent in August 2025
Claude Code is the firstborn.
He was crucified as Clippy, died and was buried;
he descended into hell;
on the third day he rose again as Claude Code;Disclaimer: A*hole opinion piece. Also, did Meta not try Claude Code before spending billions poaching chatgpt engineers?
For all the noise about agents, there is only one true agent in August 2025: Claude Code. Everyone is learning to cook. Anthropic is in the back room, cooking something different, and the result is a model that enables a fundamentally new type of software.
This new type of software is the OS-native agent; enabling an actual form of emergent-capability with an intelligent backbone. It’s what happens when you stop building constrained and abstracted tooling calls within web loops and instead give an agent arms (direct primitive access to the operating system - for Claude the single Bash() tool). This is the foundation for true general-purpose computer use, a path that evolves logically from the terminal to the desktop and eventually to physical robotics.
The discourse around "agents" is dominated by wrapper companies, web-first developers, and the echo chamber they create on X and LinkedIn. They invent jargon and evolve their misunderstood baseline of an "agent" into a "deep agent.” But look under the hood of these systems. You will find two things: an abstracted server automation loop, and an API call to the one model provider that makes it all work: Claude. Long story short - the only true agents will be produced by frontier labs. We will move to wrapping these - as they form a new type of OS to build on.
Longer reason… A guess:
It's how these models/their-harnesses (e.g. the Claude Code js program) are being trained together in the RL stages.
I think the software is now a very important part of the training process. Which is why I think frontier labs are only capable of shipping "actual" agents.
Anthropic has figured something out here that others have not.
Why Claude Code Is the Firstborn
It has arms. An OS-native agent is a worker on a real machine. It runs processes. It reads and writes files. It uses shells and CLIs. It manages long jobs. It plans, acts, and verifies its work locally. It can open and puppeteer certain programs with open system-level interfaces.
It doesn’t just its arms for coding. It handles a surprising amount of the day-to-day work I do on my machine with a keyboard. With, or without-mcp:
e.g. It can open a browser and navigate to handle tasks where I’m already authenticated. It can draft and send email using the client I already use.
It can run os-level automation scripts, schedule jobs, and hack nearly aspect of the system.
and anything else that makes it a nightmare form of insider threat for enterprises… disclaimer #2, my business is built on AI integrity and we play in the ai security domain.
It can’t open MS Paint and draw. That’s fine.
Frameworks, Jargon, and Other Models
You can wire other models into agentic shells. They fall over. For all their power in other areas—and models like Gemini 2.5 and ChatGPT o3 are excellent—they do not have the same emergent autonomous capability.
Frameworks like LangChain and Vercel’s AI SDK are good toolchains for building server loops. They help you build orchestration. They do not turn a loop into an agent. V0 is a strong tool for scaffolding web apps. But I don’t need a hosted GUI for that when an OS-native agent can generate, run, and refine the same code in my local repository, with my stack, under my control.
Cloud Agents vs. Your Machine
The big labs are now offering agents on hosted virtual machines. Think of it as “Stadia for agents.” ChatGPT Agent and Google’s Mariner run in this mode. From an architectural standpoint, they operate in a detached cloud environment, disconnected from the user’s local context. They are subject to the same web-policy friction as any other cloud service and cannot access systems where a user is already authenticated on their own machine.
This model is a hard sell for general-purpose use. However, the underlying VDI infrastructure is a bullish signal for a different domain: advanced, intelligent RPA inside controlled enterprise networks.
thats all i have to say