HOW AN AGENT WORKS
AI with a brain and arms.
The model plans the next steps. Tools carry them out. Context and memory supply the information needed for the task.
The brain: the AI model
It processes the task, develops a plan and chooses the next step. Tool feedback helps it adjust that plan.
The desk: context
This holds the information for the current step: your task, instructions, conversation history, files and tool results. The context window limits how much the model can process at once.
The notebook: memory
Stored facts and preferences can persist across sessions. The application retrieves relevant entries and supplies them to the model as context. Each product determines what is stored and retrieved.
The arms: tools
Tools read files, search the web, edit code or create calendar events. Applications connect them through APIs, browser control or MCP, for example. Each connection has its own permissions.
Autonomy is the scope for independent action
You define the goal and permissions. Within that scope, the agent can choose steps, inspect results and continue working. Approvals, budgets and stopping conditions set the boundaries.
An example: coordinating a meeting
Prepare
The agent reads authorized calendars and suggests available times. You book the meeting.
Act with approval
The agent prepares the meeting and invitation. You approve sending it.
Act independently within the task
The agent books within agreed time windows and participant groups. It pauses and reports back when conflicts arise.
How the work proceeds
- Understand the task
- Plan a step
- Use a tool
- Assess the result
The result informs the next step. The process ends at the goal, an approval point or a defined limit.
Background and sources
Terms and sources reviewed on 9 October 2026. The meeting example illustrates possible approval rules.
Agents and work environments
Desktop workspaces
Apps can combine several ways of working. For example, the ChatGPT desktop app offers general Work tasks and the Codex section; dots is a persistent agent available within it.
General-purpose agents
You describe a task in natural language. The agent handles it in its work environment, such as conducting research or preparing documents.
Coding agents
They work with source code and development tools. They can investigate errors, change features, run tests, and explain existing projects.
Agent platforms
Here you configure agents and processes yourself. You connect data sources and applications and define triggers and approvals.
Self-hosted agents
You install and operate the assistant in your own environment. You manage model access, permissions, and operations. External model APIs may also be used.
What to check before use
-
Define the result
Describe what should be available at the end and how you will know the task is complete.
-
Check access options
Clarify which files, apps, or repositories the agent needs. Grant the required access.
-
Set approvals
Choose which actions you want to review before execution, such as sending a message.
-
Estimate setup effort
Check whether an existing agent can handle the task or whether you need to configure a custom process.
-
Set cost limits
Check how usage is billed and which limits you can set.
Frequently asked questions about AI agents
Which AI agent fits my task?
Start with the desired result: research, office work, automation, or code. Then compare features, available access options, setup, and limitations. The Task filter in the catalog help with initial selection. Test a candidate with a small task of your own.
What is the difference between ChatGPT, dots, and Codex?
The shared profile view distinguishes ChatGPT with Chat and Work, dots for ongoing tasks and Codex for software work. App, CLI, and app server are access points whose requirements are listed in each profile.
How current is the information?
Each profile states the date of the sources reviewed and links to provider sources for features and limitations. Terms and availability may change afterward. The editorial method explains how the information is assessed.
Have these agents been tested first-hand?
This catalog is based on publicly documented provider information. The practical examples show possible tasks and review steps. This version does not include our own product tests or measurable performance assessments.
Start with a small task
Choose a limited task with test data and a visible result. Check how long the entire process takes, including your follow-up work. Expand its use once the process works reliably.
Choose an example promptThe profiles are based on provider documentation. This version does not include our own comparative tests or measurable performance assessments. Sources and method