LAYER 1 · PRODUCT
General AI Agent Products
ChatGPT Work · OpenClaw · Manus · WorkBuddy
A product is the user-facing system or application.
15 · General AI Agents & Computer Use
When AI Moves Beyond Chat
This module explains AI systems that can carry out multi-step work across browsers, desktop interfaces, tools, and applications.
Build a mental model of product, agent logic, capability, interaction mode, and control.
The Big Picture · four layers
LAYER 2 · AGENT LOGIC
Goal → Plan → Decide → Act → Observe → Continue
This working loop lets a system carry out multi-step tasks.
LAYER 3 · CAPABILITIES
Tool Calling · Computer Use · Browser Interaction · Desktop Interaction
Ways an agent can access or act on external systems.
LAYER 4 · CONTROL
Permissions · Human Approval · Validation · Observability
Controls determine what the agent may do and when a human must review.
CORE IDEA
A modern agent product is usually a system that combines a model, planning logic, tools, interface capabilities, state, and control mechanisms.
Products, capabilities, and behavior
A · PRODUCTS
ChatGPT Work · OpenClaw · Manus · WorkBuddy
User-facing agent products or systems. Each can package a different mix of models, tools, interfaces, and controls.
B · INTERACTION CAPABILITIES
Computer Use
AI interaction with graphical interfaces.
Browser / Desktop Agents
Agents acting in web or desktop environments.
C · AGENT BEHAVIOR
General AI Agent
A system that carries out multi-step tasks across tools and interfaces.
Tool calling is a cross-module capability; see Module 09.
What’s the difference?
Product, model, capability, and interaction method
Agent Product
A complete user-facing system that may combine a model, instructions, tools, memory or state, interfaces, permissions, and workflow logic.
Examples: ChatGPT Work, Manus, OpenClaw.
Foundation Model
A trained AI model that provides language, generation, reasoning, or multimodal capabilities.
Examples: GPT, Claude, Gemini, Qwen.
Why it matters: Model = intelligence engine. Product = complete system around it. A product may use one or more foundation models.
Computer Use
A capability that lets AI interact with graphical interfaces: click, type, scroll, select, and navigate UI.
Agent Product
A complete system that may use Computer Use as one capability among many.
Why it matters: Computer Use is something an agent can do; it is not the product itself.
Browser Agent
Usually receives a goal, observes pages, decides next actions, and adapts to intermediate results.
Browser Automation Script
Usually follows predefined rules or steps and executes expected actions with limited adaptive reasoning.
Why it matters: Both can automate a browser, but an agent typically decides dynamically while a script follows predefined logic.
Tool Calling
Requests a structured external function, such as get_order_status(order_id); a system executes it and returns structured data.
Computer Use
Interacts with a graphical interface: open browser → click Orders → search order → read the screen.
Why it matters:
| Interface | API / function | GUI |
|---|---|---|
| Input | Structured | Visual / UI |
| Typical result | Structured data | Screen / UI state |
General Agent
Designed for broad tasks: research, browser work, files, business workflows, and computer use.
Coding Agent
Specialized for software development: repositories, code editing, terminals, tests, diffs, and review.
Why it matters: A coding agent is specialized for software development; a general agent works across broader tasks. Coding Agent ⊂ Agent Products.
How these pieces can work together
One realistic workflow
USER GOAL: “Update the customer record and send a confirmation.”
GENERAL AGENT PRODUCT → PLAN
OPTION A · Tool Calling → CRM API
OPTION B · Computer Use → Browser / Desktop UI
RESULT → HUMAN APPROVAL when required → FINAL RESPONSE
Tool Calling, Computer Use, Browser Agents, and Desktop Agents are not mutually exclusive; a product can combine them.
How to Compare General Agent Systems
Browser accessDesktop accessTool integrationsFile accessComputer UsePlanningLong-running tasksHuman approvalPermissionsObservabilityModel choices
Remember the layers
A general agent product is a complete system. Computer Use, browser interaction, desktop interaction, and tool calling are capabilities the system may use to complete tasks.
The model provides intelligence; the agent system decides how that intelligence is connected to actions and controls.