What is an agent harness, and why owning one matters

A post on the Earendil blog sets out to explain, without jargon, one of the terms that has quietly become central to AI: the harness. It is written for people who see “agent harness” in every feed but have been too embarrassed to ask what it means.

Model plus harness

The simple version, the post says, is that an agent equals a model plus a harness. Where a model generates text, the harness is the software that gives it an environment to work in: instructions, tools and a control loop. Unlike the models themselves, a harness is something an end user can own and run.

Harnesses differ mainly in how you talk to them: Pi is used from a terminal, OpenClaw reaches users through iMessage or other chat apps and email, and Earendil's own harness, Lefos, was built primarily to work over email.

Four things every harness does

First, a system prompt: instructions injected with every request that govern how the model should behave in this context — less deeply embedded than a model's training, more like the briefing a new employee gets on their first day. The post points to Claude Opus 4.5's much-discussed “soul document” as a famous example of the embedded kind.

Second, tools: capabilities written in code that the model can call, such as web search, code execution or composing an email. The harness describes and provides them but does not dictate when the model should use them.

Third, the agentic loop, illustrated with an example: asked to compare local primary schools, the agent searches, judges the results insufficient, searches again on its own initiative, builds a spreadsheet with a code tool, compares it with the original request and finally composes an email with the findings attached. The loop closes when the model decides the job is done. Fourth, a translation layer that lets one harness drive models from different providers — Anthropic, OpenAI or open-weight alternatives — and even swap models inside a single loop.

Why owning the harness matters

That layer, the post argues, moves leverage from the AI labs to users. Someone who runs their own harness locally keeps the freedom to pick models on cost-per-task grounds, to modify the tool, and to hold local copies of their sessions. Pi is the example: a deliberately minimal harness with a short system prompt and few tools, free and open source, whose users have extended it to the point of sharing more than 5,000 extensions with each other.

The first widely used harness, Claude Code, was not built as a neutral layer but as an application for coding with one vendor's models. The post notes the growth since of open alternatives such as OpenClaw, OpenCode, Hermes and Pi, and says Earendil intends Pi to be neutral. The framing is political as much as technical: as concern grows about the influence of large AI companies, a harness that a person owns is, in the authors' view, one way to keep choice — and a local record of one's own correspondence with machines — in the user's hands.