What Is an Agent Harness? The User-Owned Layer That Controls AI Models

What Is an Agent Harness?

An agent harness is a software layer that provides an environment for AI models to operate as agents. Unlike the models themselves, the harness can be owned and modified by the end user. The article uses a climbing harness analogy: just as a climber owns and adjusts their harness for different routes, a user can customize their agent harness for different tasks.

Four Core Components of an Agent Harness

1. System Prompt

The system prompt is a set of instructions injected into every conversation to constrain how the AI model responds. It is similar to the "soul document" that guides models like Claude Opus 4.5, but it resides in the harness rather than being baked into the model weights. It acts like an onboarding document for a new employee.

2. Tools

Tools are code-implemented capabilities that the model can call, such as web search, code execution, or email composition. The harness describes and provides these tools but does not dictate when or how the model uses them; the model decides autonomously.

3. Agentic Loop

The agentic loop is the framework that governs the model's iterative reasoning and tool use. The article illustrates this with a concrete example: a user asks an email-based agent to compare local elementary school rankings and test scores. The agent:

Understands the request using pre-trained knowledge.

Constructs search queries and retrieves data.

Evaluates results against the original request and decides whether to search again (the first loop iteration).

Uses a code tool to create a spreadsheet for calculations and formatting.

Checks the spreadsheet against the request and may loop back for more data.

Finally calls a compose-email tool to summarize findings, attach the spreadsheet, and send the email, closing the loop.

A real Pi session demonstrating this loop is referenced (https://pi.dev/session/#b23f2459599f8439327f65c90ee95d06), though the link may be inactive.

4. Translation Layer

The translation layer allows the same harness to work with multiple AI models (e.g., Anthropic, OpenAI, open-weight models). It shifts control from AI labs to the user, enabling them to run the harness locally, keep session history on their own machine, and compare model outputs and costs side by side.

Making the Harness Your Own

Because the harness is user-owned, it can be customized. Pi is cited as a minimal, open-source harness that runs locally on the user's laptop. Users extend Pi by modifying its system prompt or writing extensions for their workflows; over 5,000 extensions have been shared among Pi users.

Neutral Open-Source Harnesses as Tools for Autonomy

Early agent harnesses like Claude Code were tied to specific vendors. Newer open-source harnesses—OpenClaw, OpenCode, Hermes, and Pi—are model-agnostic. Earendil, the maker of Pi, emphasizes building neutral harnesses and open protocols to preserve human agency, ensuring that users "swing the hammer" rather than being driven by the tool.

Royal Robbins on El Capitan, his harness loaded with climbing gear. Photo: Tom Frost (https://www.frostworksclimbing.com/cool_aid.htm)