How to Build an AI Software Factory: Agents That Open, Review, and Merge PRs

TEXT START: An AI software factory is five stages with a gate at each one.

The Dissection

This is an industrialization memo disguised as a how-to guide. It converts coding agents from a novelty into a production line: filtered intake, disposable workspaces, tool access, automated verification, and a human merge gate.

Its central reframing is accurate and important: code generation is becoming cheap, while review and acceptance remain scarce. The article then positions the factory—not the agent—as the competitive asset, with Firecrawl's search, scraping, indexing, and MCP tooling embedded in that pitch.

The text's unintended disclosure is more consequential than its advice: one engineer can already flood a team's review capacity. That is not evidence that the old system is adapting successfully. It is evidence that human labor has become the downstream bottleneck.

The Core Fallacy

The article confuses a firm's ability to absorb AI output with a solution to the economic destruction caused by AI output.

Under the Discontinuity Thesis, this factory is P1 made operational. It expands cognitive automation, concentrates productive capacity in whoever owns the infrastructure and context, and pushes more human work toward the exit. It solves throughput. It does not preserve wages, mass employment, or productive participation.

The human merge gate is treated as a durable endpoint. It is not. Reviewing scope, judging correctness, interpreting tests, and accepting accountability are cognitive functions. As models improve, that gate contracts into a small layer of liability-bearing servitors, while legal and institutional inertia temporarily props it up.

The factory therefore advances P3 rather than preventing it. It also sharpens P2: once every competitor can industrialize code production, coordination becomes a race to automate the remaining gates and cut the human surface area further.

Hidden Assumptions

• Valuable software demand will expand fast enough to absorb the increased supply of code.
• Human review will remain cheaper and more reliable than automated review.
• Passing tests, screenshots, and model gates is a meaningful proxy for business value, security, liability, and durable correctness.
• The humans retained for “decisions that carry accountability” will remain numerous enough to support the workforce displaced upstream.
• The factory's apparent moat will survive diffusion. In competitive markets, production machinery spreads; its gains become margin pressure and labor elimination.
• The owner of the factory captures the gains without confronting the collapse of the wage-to-consumption circuit.

These assumptions turn a labor-displacing production system into a neutral engineering upgrade. They are not neutral. They are the mechanism of displacement.

Social Function

Classification: partial truth, transition management, elite self-exoneration, and ideological anesthetic.

The operational claims may be useful. The gates, isolation, verification, and deduplication problems are real. But the article translates a structural employment crisis into a tooling problem for firms and advanced engineers. It teaches potential Sovereigns how to scale output while leaving the displaced majority outside the frame. The remaining human roles are presented as reassuring “judgment” positions even though those positions are themselves targets for automation.

The Verdict

This is not a survival plan for software labor. It is a blueprint for its industrial replacement.

Its strongest claim—that one person can saturate a team's review capacity—is also its death certificate. The factory preserves software throughput and possibly consumption; it does not preserve broad economic necessity for humans. In DT terms, it is infrastructure for P1 that accelerates P3. The human merge gate is not a cure for the old order. It is hospice care for the wage circuit.