What is an agent harness?

A model supplies the learned weights and generates responses. A harness is the surrounding software that decides how calls are streamed, which execution surfaces are available, how state persists, and which policies govern continuation.

Observed performance belongs to the model-and-harness combination. Changing either side can change the result, so a harness benchmark should not be presented as a property of the base model alone.

Prime Agent's harness includes its streaming agent loop, persistent IPython kernel, daemon-backed workers, append-only sessions, refinement flow, and provider adapters. Together these components turn a provider model into a stateful coding and research agent.

The architecture emphasizes programmatic execution and recoverable session state rather than a large fixed menu of model-facing JSON tools.

The harness executes generated Python, shell commands, skills, and extensions with the user's OS permissions. Process boundaries between kernels and workers improve lifecycle management but are not security confinement.

For untrusted code or instructions, isolation must come from the host environment, such as a disposable clone or an externally restricted runtime.

Harness comparisons need the same model, task, prompt, budget, version, and tool access to support a symmetric conclusion. A result obtained with different conditions cannot isolate the effect of the harness.

No harness wins every task or model pairing in the existing claims. Prefer bounded comparisons with disclosed artifacts over a universal winner label.