An agent harness is software that helps orchestrate AI agents to accomplish tasks. They generally take the form of IDE, TUI, or desktop applications.
Common features provided are:
- Model selection
- Automatic injection of an Instruction File into Context
- Slash Commands for easy Prompt reuse
- Simple inclusion of files or folders into context
- Agent Skills
- Agent selection like Plan and Build
- MCP availability to the models
- Tool Calling for enabling autonomous workflows
Integrated Development Environments (IDE) allow developers to edit code. Editors like VS Code and JetBrains IDEs provide plugins for integrating agent harnesses such as GitHub Copilot.
A new breed of IDEs have come out over the last few years that are agentic-native. These agentic first IDEs like Zed, Cursor, and Kilo provide deeper agentic coding workflows.
A Terminal User Interface (TUI) has become a popular way for companies and OSS teams to provide agent harnesses. These typically provide a chat-like interface for working with models in the terminal.
Examples of these would be applications like Claude Code, OpenCode, and Antigravity.
Big Tech has been embedding agents into operating systems like Windows and macOS for years now. More recently, Claude Cowork was launched as an example of an agentic harness with less developer focus than the harnesses discussed above.