In a chat, you ask for an explanation of a bug. You ask a coding agent to find it in a project, change files and check the fix. Both use AI, but the programs around it have different capabilities. That is where all these names come from: ChatGPT, Claude Code, Codex, OpenCode, harness, orchestrator.

Let’s arrange them into a short map: a model generates responses and decisions, an agent application organizes actions, and its interface gives you a way to control it. Then we can see why people bring several agents together.

How we arrived at familiar AI chats

A useful starting point for the current wave of interest is November 30, 2022, when OpenAI opened ChatGPT as a research preview for users to try. Conversational AI became easy to try: ask a question, request an example, follow up on an answer. Late 2022 and 2023 were a period of popularization for this way of interacting; the “boom” has no single official date.

A model is a trained system that processes its input and generates an output, such as an explanation or a proposed next action. A chat is an application through which you communicate with a model. So OpenAI, a GPT model and the ChatGPT application are different things: a company, a technological component and a user-facing product.

Anthropic is another company; Claude names its family of models and its user-facing AI assistant. Claude was introduced on March 14, 2023. Chinese developers such as the Qwen team, European developers such as Mistral and others have expanded this ecosystem well beyond two companies.

For a developer, a chat helps with understanding unfamiliar code, discussing a solution or getting a draft. But work remains between “here is a possible fix” and a fixed project: locate the files, make the changes and check their effects. Coding agents take on that part.

Why an agent needs a terminal

In this introduction, an AI agent is a program in which a model can choose subsequent actions toward a task and take their results into account. A tool is an operation available to it: read a file, search text, change code or run a check. Program code, operating with the permissions it has been given, actually executes the operation.

A terminal is a window for interacting with programs through text. CLI, or command-line interface, is a way to control a program through such an interface. Inside a coding tool you can write in ordinary language: “Find why the test fails and propose a fix.” You do not need to know every action the agent will require in advance.

Anthropic’s Claude Code and OpenAI’s Codex CLI are examples. They help investigate a project, change code and perform checks. The permitted actions depend on configuration and environment. An agent does not acquire a missing test or access merely because someone asks it to “do everything.”

By their first public CLI announcements, Claude Code came first: February 24, 2025, as a limited research preview. Codex CLI was announced on April 16, 2025. The Codex name itself is older: in 2021 it referred to a model for working with code; the modern cloud-based Codex was introduced on May 16, 2025. So “which came first?” needs a qualifier: the model, the CLI or the cloud agent.

Where the application runs also does not determine where the model runs. A CLI can run on your computer and contact a remote model through an API — a programming interface through which one program sends requests to another.

CLI, desktop and harness: different parts of the construction

A CLI fits naturally alongside development files and tools, but agents can also have graphical interfaces. OpenAI introduced the Codex desktop application on February 2, 2026. The current Claude Code documentation describes working through a terminal, code editor, desktop and browser. These are examples of products evolving, rather than a rule that every model must have its own CLI.

A desktop application is a program with windows, buttons and other graphical elements. It can present tasks and changes in a way a particular user finds more convenient. A new window does not itself give a model more knowledge or guarantee a better result.

What connects the model, files and tools? A harness is the software wrapper that supports agent work: it passes relevant data to the model, invokes tools, returns results and manages execution. This separation appears, for example, in Anthropic’s architectural explanation. Different authors may draw the boundaries of the term differently.

Context is the information a model receives for its current decision: the request, rules, retrieved passages and previous results. A capable model does not automatically see your entire computer; the wrapper needs to provide the relevant observations. Permission only to read code must remain a restriction even if the model proposes an edit.

Why OpenCode and Pi exist

You can use more than the agent application offered by a model’s developer. OpenCode is an open-source coding agent supporting multiple model providers, with terminal, desktop and web interfaces. Its purpose is to provide a development workspace where you can connect the provider you need. Having that choice does not mean every model will perform equally well.

Pi is a terminal coding agent and a collection of components for building agent applications. It is an option for people who want customization and their own extensions around the agent loop. Pi is not a separate language model; the software wrapper and model have different roles. More freedom to customize also leaves more decisions to the person doing the customizing.

Another word may appear here: a skill is a package of instructions and materials for a procedure, such as checking changes. It is not itself another agent. The same agent can follow an additional procedure when supplied with its instructions.

Branch · The wider pictureOptional · you can skip this section

OpenClaw, Hermes and a small computer

Agents also work outside software development. OpenClaw connects an AI assistant to messaging channels and tools: people can delegate information work, reminders and other automated tasks. Its capabilities depend on integrations and permissions; messaging access does not imply unlimited access to other services.

Hermes Agent, by Nous Research, is an agent system with memory, skills and operation through different interfaces, including messengers. It is intended for recurring work and retaining useful information between sessions. Distinguish Hermes Agent from models named Hermes; recorded experience alone does not establish an improvement in quality.

A Raspberry Pi is a small single-board computer, separate from the Pi coding agent. The official OpenClaw guide uses it to run a gateway — a program that receives requests and connects parts of a system — while the language model runs with a cloud provider. A small box on your desk can therefore keep an assistant working without running a large model inside it.

What happens when there are several agents?

Suppose one agent is asked to find a bug’s cause and another to check the proposed change. Several agent executors working toward a related goal make a multi-agent system. They can use the same model: their tasks, contexts and available actions also establish their differences. A delegated executor is often called a subagent.

Now someone must determine who starts first, what the reviewer receives and what happens when results disagree. Orchestration organizes this work. An orchestrator is the program or component that manages tasks, the exchange of results and completion. A coordinator can be ordinary code; another model is not always required. Anthropic’s overview describes different ways to combine these elements.

A second agent adds both possibilities and costs: it needs sufficient information, and the combined result needs checking. Two identical incorrect conclusions do not become correct because there are two agents. For a familiar small change, one executor with a good check can be sufficient.

Reference: which layer does a name belong to?

Name or conceptWhat it isTasks it servesImportant distinction
GPT, Claude modelsModelsProcess context and generate a resultA model is not the entire application.
ChatGPT, the Claude appAI assistant interfacesConverse, explain and work with informationCapabilities depend on connected features.
CLI, desktop, webWays to interactAssign a task and inspect its resultThe interface does not determine model intelligence.
Claude Code, CodexCoding toolsInvestigate projects, make and check changesSpecify the product, interface and environment.
OpenCode, PiOpen coding agents and componentsWork with code and customize an agent workspaceThey are wrappers that use models.
HarnessAn agent’s software wrapperConnect context, a model and tool executionPermissions and checks have a concrete implementation.
Multi-agent systemSeveral agent executorsDivide and check related workMore executors do not guarantee quality.
OrchestratorA control componentCoordinate tasks, dependencies and completionIt can work without an AI model of its own.
On a narrow screen, scroll the table sideways. With a keyboard: Tab to the table, then use the left/right arrow keys.

Where a beginner can start

  1. Need to understand code or discuss a solution? Start with a chat. Ask it to explain assumptions and check important conclusions; the familiar conversational interface already provides a useful starting point.
  2. Need to work on a project? Try one coding agent on a small task. Choose an interface you understand and check support for your system. Inspect which files changed and what was actually checked before trusting it with larger work.
  3. Add agents for a specific need. For example, a separate check or independent subtasks, with clear roles and a way to assemble the result. Getting acquainted with these tools does not require starting with a large “agent factory.”

Next time you encounter a new name, first place it on the map: is it a model, a program around one, an interface or a way to organize several executors? That makes it easier to understand the tool’s use and see what its name alone does not promise.

Announcement dates and the capabilities described were checked against primary sources on November 17, 2026. Tools change; this is an introductory map, not a comparative test of their quality.