Back to Learn

What is an Agent Harness?

A harness is the runtime environment that executes AI agents—the software that connects a language model to tools, files, and the outside world.

The Agent Runtime

Think of a harness as the "operating system" for an AI agent. It provides:

  • Long-running execution

    Agents can work for extended periods, handling multi-step tasks that span minutes or hours.

  • Tool integration

    Access to file systems, terminals, browsers, APIs, and other capabilities the agent needs.

  • Skill loading

    Skills get "attached" to the harness, extending what the agent can do.

How Harnesses Work

1

Startup

The harness initializes, scans for available skills, and establishes connections to tools (file system, terminal, etc.).

2

Request Handling

User sends a task. The harness prepares context (relevant files, skill descriptions) and sends it to the language model.

3

Agentic Loop

The model responds with actions (tool calls). The harness executes them, feeds results back, and continues until the task is complete.

4

Skill Invocation

When the model identifies a relevant skill, it loads the full skill content and follows those instructions for the task.

Skills Attach to Harnesses

Skills are not standalone applications—they need a harness to run. When you install a skill:

  • The harness discovers the skill in its configured skill directories
  • It reads the skill's name and description (lightweight metadata)
  • When a task matches, it loads the full skill content into context
  • The agent follows the skill's instructions using the harness's tools

Example Harnesses

Different vendors provide harnesses with varying capabilities:

Claude Code(Anthropic)Cursor(Cursor)Goose(Block)Amp(Sourcegraph)OpenCode(OpenCode)GitHub Copilot(GitHub)OpenAI Codex(OpenAI)Letta(Letta)

Why Harness Choice Matters

Tool Availability

Some harnesses have browser automation, others don't. Some can execute arbitrary code, others are sandboxed.

Model Options

Each harness may support different language models—Claude, GPT-4, Gemini, or specialized coding models.

Execution Environment

CLI vs. IDE integration vs. web-based. Local execution vs. cloud-hosted. Interactive vs. batch mode.

Skill Compatibility

Not all skills work on all harnesses. A skill requiring shell access won't work on a sandboxed harness.

UNKNOWN ENV