The Model is a Commodity. The Harness is the Advantage.

The article, based on a translation of Eyad Khrais's writing on harness engineering, posits that AI models are commodities accessed via API, while the harness — the deterministic tools and context — is where value is built. The author shares personal experience from a finance-focused project: prompt engineering is fading, replaced by context engineering (coined by Tobi Lütke and Andrej Karpathy in summer 2025). A thin prompt combined with thick, well-curated context and deterministic CLI tools yields predictable results. The article warns against "context rot" (quality degradation with long inputs, as shown by Chroma's tests) and emphasizes that deterministic code (e.g., CLI commands) prevents hallucinations and ensures accuracy, especially for critical tasks. The harness constitutes about a third of the project's complexity but pays off by improving reliability. The author provides rules for creating effective tools (narrow, deterministic, understandable to the model, grouped, with validation) and advises that reusable tasks should be turned into CLI commands. While acknowledging the "Bitter Lesson" that handcrafted structures may be eroded by more powerful models, the author argues that coordination, state, error recovery, and access rights will remain essential harness functions. The boundary: the model handles language, judgment, and planning; the harness handles counting, rules, state, and external actions. Ultimately, the role of "prompt engineer" is disappearing, replaced by builders of project-specific harness, memory, and context.