AI coding agents, compared.

What they are, how they differ, and who should check the work.

An AI coding agent is a program that lets an AI model work on your files. You give it a goal. It reads the code, makes changes, runs commands and checks its own work. The program around the model is called a harness. Choose one by where it runs, who makes it and which models it can use. A person should still review every change.

Facts checked on November 18, 2026.

Words on this page

12 short definitions.
model
The part of an AI tool that uses what it learned to give an answer. Claude and GPT are model families.
provider
A company that offers AI models you can use, such as OpenAI, Anthropic or Google.
harness
The program around a model. It lets the model open files, run commands and use tools.
coding agent
An AI tool that works on code for you. You give it a goal, and it makes the changes.
terminal
A text window where you type commands to your computer. It is also called the command line.
IDE
The program where people write code, also called a code editor. VS Code and the JetBrains editors are well-known ones.
open source
The code is public and its license lets anyone use, change and share it under stated rules.
license
The legal rules for who can use, copy or change the code. MIT and Apache-2.0 are two common open source licenses. Proprietary code can have tighter rules about who can use or change it.
local model
An AI model that runs on your own computer instead of on a company's servers.
SDK
A kit that lets developers build their own tools on top of another tool.
cloud agent
An agent that runs on remote computers you reach over the internet.
developer preview
An early version that the maker has released for developers to try.

What a harness is

A model can suggest a change to your code. A harness is the program that lets the model make that change in your files. It lets the model open files, run commands and use tools. A harness can also ask you before it does something risky, and some let you choose how much it asks.

Think of the model as an engine and the harness as the car. The engine makes the power. The car has the steering, the brakes and the dashboard. Two cars can use the same engine and still drive very differently.

That is why the same model can feel different in different tools. The harness decides what the model can touch, what it remembers, and when it has to stop and ask.

The tools, side by side

Here are thirteen AI agent tools. Some are built for code and some do other work on a computer, like OpenClaw. This is not a full list of every tool.

The source links go to the makers' own pages and code. If a maker has not published a fact, the table says so instead of guessing.

AI coding agents compared. Checked on 2026-11-18.
ToolMade byRuns inOpen sourceModelsSources
Claude CodeAnthropicTerminal, VS Code, JetBrains, Desktop app, Web, PhoneNo (Proprietary, all rights kept by Anthropic)Claude models from Anthropic or through other providers.
CodexOpenAITerminal, IDE extension, ChatGPT desktop app, ChatGPT mobile app, ChatGPT on the web, Codex CloudYes (Apache-2.0, the command line tool)OpenAI models by default. It can also use other providers, and local models.
CursorAnysphere, Inc.Desktop app, Terminal, Cloud agent, Web, MobileNo (App rights reserved)Models from OpenAI, Anthropic, Google and more, plus its own. Access varies by plan.
OpenCodeanomalycoTerminal, Desktop appYes (MIT)Models that run on remote servers or on your own computer.
Hermes AgentNous ResearchTerminal, Desktop app, Messaging apps, Hosted cloudYes (MIT)Many providers, including local models.
PiEarendil Inc.Terminal, Your own code (SDK)Yes (MIT)Many providers, including local models.
ClineCline Bot Inc.VS Code, JetBrains, Terminal, Desktop app, SDKYes (Apache-2.0, excludes the JetBrains plugin)Many providers. It is not tied to one, and it works with local models.
Grok BuildSpaceXSITerminalYes (Apache-2.0)Supports models that run on your own computer.
DeepSeek HarnessIn developer previewDeepSeek AIYour own computer, in a web page that opens thereYes (MIT)DeepSeek and other providers.
MiniMax CodeMiniMaxTerminalYes (MIT, terminal version)A MiniMax account, or your own provider.
OpenClawThe OpenClaw Foundation and its communityYour own computer, Messaging appsYes (MIT)Models that run on remote servers or on your own computer.
Atomic AgentAtomicBot-aimacOS, Linux, WindowsYes (MIT)Local models by default. It can also connect to cloud model providers.
Gemini CLIGoogleTerminalYes (Apache-2.0)Gemini models.

Checked on November 18, 2026. Not published means the maker has not said so where we looked.

How to choose one

Start with where you want it to run. Some tools live in the terminal. Some work inside an editor. Some run in the cloud, on the maker's servers, instead of on your computer.

Next, look at who makes it and whether the code is open. Open source means you can use, change and share the code under its license. Check which parts of each tool the license covers.

Then check which models it can use. Some tools only work with one company's models. Others let you switch between many, including models that run on your own computer.

Last, decide who reviews the work. Every tool here can make mistakes. Pick the one that makes review easy for your team, and keep a person in charge of what gets approved.

How this connects to what we build

Coding agents work on software. The agents we build for teams do business work, like follow-ups, scheduling and data entry. The idea is the same in both. The AI does the task, and a person approves what matters.

Questions people ask

What is an AI coding agent?

It is a program that lets an AI model work on your code. It can read files, change them and run commands, then report back. You set the goal and review the result.

What is the difference between a coding agent and a chatbot?

A chatbot answers questions in a chat window. A coding agent can also open your files, change them and run commands. That is why it needs permission settings and a person who reviews its work.

Do I need to know how to code to use one?

You need enough skill to judge the result. The agent writes the code, but someone still has to check that it does what you asked and does not break anything. If nobody on your team can review code, ask for help before you start.

Are open source coding agents safer?

Open source means you can read how the tool works, which helps you check what it does. It does not make the model's answers correct. Safety comes from limiting what the tool can touch and from reviewing every change.

Which models do coding agents use?

It depends on the tool. Some only work with one company's models. Others connect to many providers, and some can use models that run on your own computer. The table shows what each maker says.

How current is this page?

Each row links to the maker's own sources. The date shows at the top of the page and under the table. If a fact is missing, the maker has not published it where we looked.

Keep reading