Agentic Harness Tools
Definition
A harness tool is a system used to run, evaluate, and compare AI models in a consistent way.
The key distinction
| Concept | What it is | Analogy |
|---|---|---|
| Model | The LLM which receives tokens and generates tokens. No memory, tools, or interface on its own. | The engine |
| Harness | The system around the model which has memory, tools, prompts, UI, workflows, and orchestration. | The car |
Exmaple: The harness is why Copilot in VS Code feels different from Claude Code in the terminal, even when both run the same underlying model.
The main agentic harnesses compared
GitHub Copilot:
- Available as a CLI tool and as editor extension for VS Code, IntelliJ, and other IDEs
- Developed by GitHub
Claude Code:
- Terminal-based AI coding agent focused on large codebases and multi-file tasks
- Developed by Anthropic
Codex CLI:
- Terminal-based AI coding agent for repository-aware coding and automation
- Developed by OpenAI
OpenCode:
- Terminal-based AI coding agent which can connect all your subscriptions
- Open-source
Cursor:
- AI-first code editor with built-in chat, autocomplete, and agent workflows
- Developed by Cursor
What the harness controls
| Harness | Description |
|---|---|
| System prompt design | Sets baseline rules and behavior before user input, shaping tone and reasoning |
| Context selection | Controls what data (files, history, docs) the model can see |
| Tool availability | Defines what actions the agent can perform beyond text output |
| Agent loop design | Controls how the agent plans, retries, and iterates on tasks |
| Compaction strategy | Determines what gets summarized or removed when context is full |
Most AI coding harness tools often rely on the same underlying models, but the outputs can still feel quite different. This is because each tool wraps the model in its own system.