AI agents compared
An AI agent does not stop at an answer: it clicks, types, runs commands and finishes the job. There are now many of them, and they differ in two ways that matter most: where they work, and who says yes before something important happens. Here is a fair overview (September 2026).
Agents for developers
Claude Code, from Anthropic, works in your terminal and on your files: it reads a codebase, makes changes and runs tests, asking before commands you have not allowed. Claude can also use a computer, seeing the screen and moving the mouse. OpenClaw is the open source option: you run it yourself with your own AI key, it works around the clock, and you reach it through WhatsApp, Telegram or Discord.
These are harnesses: a model plus tools, permissions and memory. They are powerful and flexible, and they expect you to set them up and decide what they may do.
Agents in the cloud
Manus and OpenAI's agent features in ChatGPT take a goal and work on a computer in the cloud, browsing sites and producing documents, slides or code. They shine on large, self-contained tasks. Because they work in their own browser, signing in to your accounts and paying are the hard part, and names and features change often.
Agents in the browser
Perplexity's Comet builds the agent into a web browser, free for everyone: it summarises pages, compares tabs and can act on a site for you. It is the easiest start if your work happens in the browser.
A personal agent: Iris
Iris is an agent for your personal life, ready to use. Each account gets its own sealed-off assistant with its own computer, memory and files, hosted in Europe. It works in your mail, calendar and messages, on your iPhone, Mac, tv and in Chrome, and you can talk to it like a person.
Its safety model is simple: every power is a switch that starts off, every action that counts waits for your Face ID or Touch ID, passwords stay in a vault the assistant never sees, and everything is written to a log you can read.
Be among the first
Have an invite? Log in. Using Iris means you accept the terms of service. The privacy statement explains what we keep.
Questions
What is an agent harness?
The software around a language model that lets it act: the tools it may use, the permissions it needs, its memory and the loop that keeps it working until the task is done. Claude Code and OpenClaw are harnesses; Iris is one too, packaged as a personal assistant.
Which agent is the safest?
Look at three things: where your data lives, whether it asks before actions that count, and how it handles passwords. Iris asks with Face ID or Touch ID for each important action and keeps secrets in a vault the model never reads.
Which one should I pick?
For code, Claude Code. For big research or build jobs, a cloud agent such as Manus. For the web, Comet. For your own mail, agenda and errands across phone, Mac and tv, Iris.