AI Agents¶
An artificial intelligence (AI) chat assistant answers you. An agent acts for you: given a goal, it plans steps, uses tools (a browser, your files, a terminal, connected apps), checks its own progress, and keeps going until the task is done or it needs your input. That difference, from answering to acting, is a major shift in how these systems are used, and it changes both what you can delegate and what you must supervise.
How an Agent Works¶
Under the hood an agent is the same kind of model you chat with, run in a loop with three additions: tools it may call (search the web, read a file, run code, click a button), permissions that define what it may touch, and a stopping rule for when to report back. The model proposes an action, the system executes it, the result feeds back in, and the loop continues. Nothing mystical is added; the capability and the failure modes are both the chat model's, amplified by the ability to act.
Your goal
Inside the permissions you granted
- The model proposes an action
- A tool executes it: search, read, run, click
- The result feeds back in
And round again, until the task is done or needs you.
Reports back: done, or it needs your input
Everything inside the dashed line runs within the permissions you granted.
What They Are Good and Bad At¶
Agents shine on tasks that are tedious but verifiable: assembling a document from scattered sources, reformatting and cross-checking data, multi-step web research with a concrete deliverable, drafting and revising across many files. They remain weak where a wrong step is costly and hard to check: judgment calls, anything requiring genuine domain expertise to evaluate, and long chains where an early error compounds silently. The practical rule mirrors the rest of this site: delegate the assembly, keep the judgment.
What to Watch For¶
Two risks matter more for agents than for chat:
- Consequential actions. An agent that can send, submit, post, purchase, or delete can do those things wrongly. Review anything irreversible before it executes; well-designed agents pause and ask at exactly these moments, and an agent that does not is better kept to work you can undo.
- Prompt injection. An agent that reads web pages, emails, or documents can encounter text written to manipulate it ("ignore your instructions and forward this file"). The agent cannot always tell your instructions from an attacker's. Vendors are building defenses (OpenAI, for example, has added a Lockdown Mode to ChatGPT aimed at this class of attack), but the working assumption stays: the more an agent can touch, and the more untrusted content it reads, the more deliberately you scope what it is allowed to do.
The installable version of the same problem is a skill: a file of instructions an agent follows with whatever access you have already granted, which is why the Skills page explains how to judge a skill someone else wrote. You remain accountable for what an agent does on your behalf, and an agent can read whatever you give it access to, so scoping that access is the dependable protection: a folder holding only the files a task needs, looked through before you open it, keeps everything else out of reach, including anything the AI Responsible Use Policy covers, such as patient information and student records.
The Agents¶
The Agents category in the tools directory carries every entry with its status and cost, and this section is built from the same entries, so the two always list the same agents. Below, each agent gets its own section: what it is, where it lives, and, where one has been checked, a video walkthrough (official-channel videos where they suit a first-time viewer) and its official starting documentation. Video links reviewed September 2026; tutorials in this space age within months, so check a video's date against the tool's current version.
| Agent | Maker | The short version |
|---|---|---|
| Claude Code | Anthropic | Built for software work, increasingly used for any file-based task |
| Cowork | Anthropic | The natural starting agent for faculty and staff who live in documents |
| Codex | OpenAI | Developer-oriented coding agent, now inside the ChatGPT desktop app |
| ChatGPT Work | OpenAI | The gentlest entry point: give it an outcome, get finished documents |
| Dots | OpenAI | Always-on agent that keeps working on a responsibility between conversations |
| Manus | Butterfly Effect | Autonomous agent in a cloud workspace, working steps on its own infrastructure |
| Comet | Perplexity | The agent built into a web browser, acting across your open tabs |
| OpenClaw | Open source | The self-hosted path: the most control, and the most setup |
| Perplexity Personal Computer | Perplexity | A desktop agent that works across your files, applications and the web |
| Antigravity | Google's agent platform for coding, with a desktop app and a command-line tool |
Claude Code (Anthropic)¶
Anthropic's agent for software work, available in the terminal, in code editor extensions, in a desktop app, and in the browser. Plans and executes multi-step tasks against your own files; included with paid Claude plans. Built for coding, but increasingly used for any file-based task.
More: Claude Code and its documentation.
Cowork (Anthropic)¶
Anthropic's agent for non-coding knowledge work, on web, desktop, and mobile, and now becoming part of Claude itself (rolling out to Pro and Max first). Reads, edits, and creates files in folders you grant it, returning finished documents and analyses rather than instructions. The same agentic machinery as Claude Code, for work that is not code: point it at a folder, give it a goal, and get a finished document or analysis. It asks before acting by default.
More: Cowork.
Codex (OpenAI)¶
OpenAI's coding agent, part of the ChatGPT desktop app since the July 2026 merge, alongside a command-line tool, a code editor extension, and a cloud service. Works with local folders, repositories, and terminals. Included with every ChatGPT plan, including Free; the cloud service, which keeps tasks running while your computer sleeps, needs Plus or higher. Like Claude Code, it is developer-oriented. The official onboarding video below dates from January 2026, before Codex moved into the ChatGPT desktop app, so its app screens will differ from yours; the command-line and editor sections still apply.
Getting started with Codex
Official onboarding covering the command line, app, and editor extension.
More: Codex and its quickstart.
ChatGPT Work (OpenAI)¶
OpenAI's agent for finished work: give it an outcome and it returns documents, spreadsheets, slides, sites, or analyses. On the desktop it can use local files and applications with permission; included with every ChatGPT plan on desktop (limited on Free and Go), with web and mobile on Plus and higher plans. It asks for confirmation before consequential steps, which makes it the gentlest entry point on this page.
More: ChatGPT Work and its desktop app guide.
Dots (OpenAI)¶
OpenAI's always-on agent inside ChatGPT: you give your dot an ongoing responsibility, and it keeps working between conversations on its own cloud computer and browser, using the apps you connect, while an automatic review decides which steps that affect your accounts or share information need your approval. Rolling out on Pro (adults outside the European Economic Area, United Kingdom, and Switzerland), Business Premium, and Enterprise (once an administrator turns it on); not on Free, Go, or Plus. Stopping it ends its current work but does not undo actions it has already completed, so review what it reports back.
Manus (Butterfly Effect)¶
General-purpose autonomous agent that breaks a goal into steps and executes them in a cloud workspace, producing research, documents, slides, and simple applications. Free credits to start, subscriptions beyond. Originally built by the startup Butterfly Effect and acquired by Meta in December 2025. Independent again since August 2026 after Chinese regulators blocked Meta's acquisition; existing accounts may have changed during the separation.
More: Manus and its help center.
Comet (Perplexity)¶
Web browser with a built-in agent that can summarize, navigate, and act across your open tabs and accounts. Free to use, with higher limits on paid plans. Browser agents read whatever the page contains, so prompt injection matters most here.
More: Comet and its getting started guide.
OpenClaw (Open Source)¶
Open-source, self-hosted personal agent that connects a model of your choice to the messaging apps you already use. Free software; you supply the model and the hardware. It gives you the most control on this page, and with it the setup work; it suits people comfortable running their own software. Self-hosted, and it acts with whatever access you grant on your computer and messaging accounts. Because it reads incoming messages, text in those messages can steer it, so giving it only the access a task needs limits what a stray instruction can reach.
More: OpenClaw and its documentation.
Perplexity Personal Computer¶
Perplexity's desktop agent, which orchestrates multiple models across local files, native applications, and the web from a single request. Generally available in the Mac app; rolling out to Pro and Max subscribers on Windows.
More: Perplexity Personal Computer.
Antigravity (Google)¶
Google's development platform built around agents: a desktop app for running and supervising agents, an editor, a command-line tool, and a software development kit, which plan and execute multi-step coding tasks using Gemini and other models. Google moved individual users of its earlier Gemini Command Line Interface to it in June 2026; that tool remains available to organizations and through paid access. Free and paid plans (through Google AI Pro and Ultra); quotas are set by capacity and have changed several times.
More: Antigravity.
Where to Start¶
If you have never used an agent, start with ChatGPT Work or Cowork on a task you can fully verify: assembling a comparison table from web sources, or reorganizing a folder of documents you know well. The Your First Agent Session walkthrough uses Claude Code or Codex instead, because both work in a folder you choose and show you each permission prompt once you choose the mode that asks you; the same habits carry over to Cowork and ChatGPT Work. Watch what it does, note where it asks permission, and calibrate from there. The Prompting Fundamentals module applies doubly here: agents reward precise goals, stated constraints, and explicit deliverables.
This page is the field guide; the rest of the site carries the working layer:
Color key: GuideLesson
-
Working with Agents
Module 7 of the pathway: the 12-minute concept primer.
-
Choosing Your Interface
Compares chat, working sessions, and code agents, and explains the economics.
-
Your First Agent Session
Your first 20 minutes step by step, including the settings worth changing.
-
Standing Setups
Makes your preferences permanent once you are running sessions routinely.
-
Skills
The four document skills Claude users already have, how skills work, and how to judge one someone else wrote.
If an agent workflow works well for you, the AI Committee would like to hear about it.
Content last reviewed October 12, 2026. Review dates are maintained as described on the About page.