Implementation That Favors the AI, Review That Favors the Human---Writing a COW Filesystem from Scratch, Part 2

Description and use

The author implemented a multi‑party verification process where a Main Agent delegates to Strong, Supporter, and Opponent agents to reduce hallucinations and bias in designing a COW filesystem.

Problem

Designing a copy‑on‑write filesystem from scratch with an AI agent requires a clear blueprint and a robust verification process to avoid hallucinations and biased conclusions.

Outcome

The author achieved a clearer, more reliable design process for the COW filesystem, reducing reliance on a single AI and enabling focused human oversight.

Workflow

• Define design decisions (e.g., number of rooms, data format).
• Main Agent assigns Strong Agent to gather information and propose an answer.
• Supporter Agent searches for independent supporting evidence.
• Opponent Agent searches for strong counter‑evidence.
• Main Agent verifies validity of all reports and, if valid, repeats the process for a second round. If no result after three rounds, hand to human.
• Human reviews final summary and makes trade‑off decisions.

Evidence

The implementer is an AI Agent. An Agent can write code toward a goal, but we lack a blueprint.

Source

The conversation where the task starts is called the Main Agent. It only asks questions according to the plan.

Source

Supporter B looks for non-overlapping supporting evidence for the answer, that is, other arguments that also support it.

Source

Opponent C looks for arguments that show the answer does not hold at all. If he finds any, he reports them; otherwise he reports that he found none.

Source

If, in this era, AI context has reached the million-token scale, the context window can even hold all the code and documentation of a medium-sized project at once.

Source