AI agents you can use today, sorted by job

“Agent” now covers almost anything with a model inside it. The autocomplete in your editor is sold as one, and so is the assistant that opens a browser, signs in to your accounts and books the flight. Asking which agent is best is like asking which vehicle is best. I have spent two years building fleet software, and the first question there is always what you are carrying.

So this is a library, shelved by the job rather than the brand. For each agent: what it is good at, where I would put it to work and what I would keep an eye on. It is a snapshot of the start of October 2026, and the shelves move fast. OpenAI alone shipped three different kinds of agent inside ChatGPT this year, and in June it announced that Agent Builder, the visual canvas it launched in October 2025, will stop working on 30 November. Choose the job first. The product name is the part most likely to change.

If you want agents to build rather than to use, the companion piece is a guide to the agent frameworks you can import like a library.

Two questions to ask of any of them

First, what can it touch? An agent that drafts is making a suggestion. One that sends, merges or pays is taking an action, and those deserve different permissions. Second, what happens the tenth time it gets something wrong: who notices, and how do they undo it? I have written about this as the alarm problem, and every entry below is easier to judge with both questions in mind.

AI coding agents

The most crowded shelf, and the most mature. The honest way to divide it is by where you want to supervise the work: in the terminal, in the editor or in a pull request.

Claude Code

Anthropic's agent began in the terminal and now runs in editors, the desktop app and the browser too. It reads the repository, runs commands, edits files and checks its work against your tests, and it is strongest on a change that spans many files in a codebase it can explore. Watch the permissions. It acts with whatever access the person running it has, so the list of commands you allow is a product decision, not a setting. Its loop is also published as a library, which the framework guide covers.

OpenAI Codex

OpenAI's coding agent works in the cloud as well as in a terminal and an editor. Its natural rhythm is the background job: you hand over a task, it works in a sandbox and comes back with a change to review. That suits a backlog of well-described tickets far better than a design problem nobody has written down yet. The Codex CLI is open source, if you want to see how the loop is put together.

GitHub Copilot

If your team already lives in GitHub, Copilot's coding agent removes a step rather than adding a tool. Assign it an issue and it opens a pull request, which goes through the same review as anyone else's. That makes for a duller demo and a better fit for an organisation, because nothing about the workflow changes except who wrote the first draft.

Cursor

An editor built around the model, for people who want to watch each step and steer as it goes. If your instinct is to read a diff while it is being written rather than afterwards, start here.

The comparisons published this year disagree about which of these writes the best code, and the disagreements tend to follow whoever ran the test. What does not change is the bill on the other side: the generated code still has to be reviewed, and review is where the time goes now.

AI agents for office work

This is the shelf most people outside engineering will actually use. These agents work through documents, spreadsheets, inboxes and folders, and come back with something finished rather than a reply.

ChatGPT: Work, workspace agents and dots

OpenAI now has three shapes of agent inside ChatGPT, launched within six months of each other. Workspace agents (April) are shared agents a team builds once, connects to tools such as Slack and Salesforce and lets run in the cloud, with approval steps where they matter; for companies they replace custom GPTs. ChatGPT Work (9 July) takes an outcome rather than a prompt and works for hours towards a finished spreadsheet, deck or document. Dots, announced at DevDay on 29 September and still rolling out, are always-on agents with a computer of their own in the cloud, which you can message from ChatGPT, Slack or Teams.

Which one to reach for follows from the job. A recurring team process is a workspace agent. A one-off piece of work with a clear finish line is Work. Something you want watched continuously is what dots are for, and it is the one I would introduce most carefully, because an agent that never stops is an alarm that never stops.

Claude

Anthropic took the Claude Code idea to people who will never open a terminal. It started in January as Cowork, a separate mode where you handed Claude a folder and a multi-step job. Since 16 September it has simply been part of the Claude app: you describe what you want, and Claude decides whether that needs an answer or a piece of work. It can work on your files and inside connected apps such as Gmail, Google Drive and Slack, and the merged version is reaching Pro and Max subscribers first.

Microsoft 365 Copilot and Gemini Enterprise

If your company runs on Microsoft 365 or Google Workspace, the agent already inside your tenancy has one advantage none of the others can match: it has passed your security review. A slightly better agent that needs six months of procurement is, for most teams, a worse agent.

AI browser agents

These drive a web browser the way you would: click, type, scroll, fill in the form. ChatGPT's agent mode does it inside ChatGPT (OpenAI has been folding its Atlas browser back into ChatGPT itself), Anthropic's Claude in Chrome does it from an extension, and Perplexity built an entire browser, Comet, around the idea.

They are the most impressive to watch and the easiest to get into trouble with. A browser agent reads every page it visits, and a page can carry instructions written for the agent rather than for you. That is prompt injection, and no vendor claims to have solved it. My rule here is simple: begin with tasks that only read, keep it away from anything that pays, and sign it out of every account the task does not need.

AI research agents

The deep research modes in ChatGPT, Gemini, Claude and Perplexity read dozens of sources and come back with a cited report. For a first pass at a market, a competitor set or a technical question, they turn an afternoon of open tabs into a few minutes of waiting.

Read the citations, though. These tools sound equally sure whether a source says the thing or merely sits near it. Treat the report as a reading list with a draft attached, not as an answer.

AI agents for customer support

Support is where agents have been in production longest, because the job is narrow and the outcome is countable. Intercom's Fin, Sierra and Decagon are built for it, and Salesforce's Agentforce is the natural pick if your support already runs on Salesforce.

Vendors will quote a deflection rate. Ask for the resolution rate instead, and ask what happens at the handoff. A conversation the agent cannot finish should reach a person with the history attached, the way a fleet alarm should become a ticket with an owner. A customer who is deflected here and comes back angry through another channel has not been helped. They have been moved.

AI agents you host yourself

OpenClaw

The open-source personal agent that took over developer feeds this year. It runs on your own machine or server, lives in the messaging apps you already use (WhatsApp, Telegram, Slack, Discord) and works on a schedule, watching sites, sorting email and running scripts while you sleep. OpenClaw is MIT-licensed and governed by a foundation; its creator joined OpenAI in February, and the project stayed independent.

It is also the clearest example of the trade every agent here asks of you. It can do a great deal because you give it a great deal: your shell and your files, and often your accounts. Its skills are folders of instructions and scripts that anyone can publish, so installing one means running somebody else's code. Keep it on a machine that holds nothing you would mind losing until you have read what you installed.

n8n

When the job is a pipeline rather than a conversation, a workflow tool with agent steps is often the better build. The n8n workflow tool can be self-hosted, and its agent node lets a model make the decisions inside a flow you can see, test and version. The flow is the part that survives when you swap the model.

How I would choose

Start from where the work ends up and who signs it off. Code ends in a pull request, so choose an agent that delivers to one. A report ends with a manager, so choose the one that lives where that manager already reads.

Then set the autonomy per action, not per product. A draft is cheap to undo. A sent email is not, and a payment is worse. Most of these tools now let you require approval for some actions and not others; use that, and loosen it only when you have evidence that it is safe to.

And keep your work portable. The instructions you give an agent can live in plain files: Agent Skills, the SKILL.md format Anthropic released as an open standard, is now read by Claude Code, Codex, Copilot and Cursor alike. Your connections to other systems can go through the Model Context Protocol, which the Linux Foundation now looks after. Agent Builder's users have until 30 November to rebuild, and the ones who kept their logic outside the canvas will have the easier month.

What I would measure

Not how often people open it. How many of its actions a person reversed. How many of its drafts went out without an edit. How long a task took with it compared with the same task without it, measured rather than felt. And the number that tells you whether trust is growing or quietly leaking away: how many permissions people still allow it after three months, compared with the first week.