The Brain vs. The Jar

AI models (like GPT-4 or Claude) are essentially 'brains in a jar.' While they are highly capable, they are inherently limited by their inability to interact with the outside world—they cannot natively run code, browse the web, or access files. The performance gap between AI products often isn't due to the underlying model, as benchmark scores between top labs have converged. Instead, the difference lies in the agentic harness: the architectural wrapper that enables the model to act, remember, and iterate.

The Three Pillars of the Agentic Harness

To transform a static model into a functional agent, the harness must provide three critical capabilities:

• Tools: These allow the model to interface with the world. Beyond basic file I/O, modern harnesses enable code execution in sandboxes, web browsing, and 'computer use' (controlling a cursor to operate software). The Model Context Protocol (MCP) is emerging as a standard to allow these tools to plug into various harnesses without custom rebuilds.
• Memory: Because models have finite context windows that reset after a session, the harness must manage persistence. This includes loading instruction files (e.g., agents.md) to define project conventions, compacting context windows by summarizing past interactions, and implementing retrieval systems (semantic search or code indexing) to feed the model only the relevant data rather than the entire codebase.
• Agentic Loops: This is the operational cycle where the model plans a step, the harness executes it, and the model observes the result. Sophisticated harnesses add a verification layer to this loop, such as running automated tests, taking screenshots of UI states, or using a secondary 'reviewer' model to ensure the agent stays on track during long-running tasks.

The Fluid Future of AI Architecture

It is a mistake to evaluate AI capability based solely on the model. A brilliant model can fail in a poorly designed harness, while a mid-tier model can perform exceptionally well in a robust one. The distinction between the two is becoming increasingly fluid: complex behaviors like long-horizon planning are being trained directly into models, while consistency and project-specific knowledge are increasingly managed by harness-level conventions. Builders should focus on optimizing the 'jar' as much as the 'brain' to achieve production-grade results.