The emergence of "Reverse Engineer Anything" (REA), a meta-command-processor (MCP), is generating significant discussion within the software development community. REA purports to enable users to virtually rebuild any software by decompiling its source application and then instructing an AI agent to recreate or replicate its features. Demonstrations have included reverse engineering an Apple TV remote in approximately 45 minutes, and claims have been made about reproducing the entire suite of Adobe Photoshop applications. This article examines the functionality of REA, critiques common interpretations of its impact, and discusses its broader implications for the future of software.

How Reverse Engineer Anything Works

REA operates by facilitating communication between a human user and an AI agent. The user poses a query, such as "How does this feature work?" or "Can you make me this app?", to their AI agent. The agent then transmits this request to the REA MCP. REA, in turn, interfaces with a suite of established reverse engineering tools, including Ghidra (released in 2019), Hopper (released in 2011), and IDA (released in the 1990s). This integration allows the AI agent to interact seamlessly with these tools. REA also leverages services like Chrome DevTools to further analyze applications.

The output typically includes pseudocode, which provides high-level descriptions of functionality, proof of operation, and identification of unknown elements. The AI agent then iteratively processes this information, refining its understanding and rebuilding the application.

Setting up REA is described as straightforward. Users can instruct their AI agent to download and install REA. The process can be managed within terminal applications like Orca, which allows for the creation of multiple terminal subdivisions to run different AI models concurrently, such as Claude and Codex.

A Practical Demonstration: Rebuilding the Preview App

To illustrate REA's capabilities, a demonstration involved rebuilding the macOS Preview application. Preview, a standard photo viewer, offers functionalities such as information display, zoom in/out, and sharing. The objective was to replicate these features, potentially with added capabilities like drawing boxes or handwriting annotations.

The AI agent was instructed to "Use REA to rebuild the preview app on my Mac." Concurrently, a separate attempt was made to rebuild the same app without using REA, instructing an AI agent to "reproduce the Mac preview app, the photo viewer one, and then serve me the result."

The REA process involved decompilation using Ghidra, with the AI agent progressing through the rebuilding stages. The non-REA attempt also proceeded independently.

The first completed result was from Opus, without relying on REA. The resulting Preview app exhibited some lag, potentially due to background processes. The REA-developed version took an additional 3-4 minutes to complete. Upon testing, the REA version opened and closed various tabs and displayed flashing elements. It eventually launched, presenting three files.

A comparison between the REA-developed app and the original macOS Preview app revealed significant similarities. Both applications displayed titles, information panels, zoom functionality, sharing options, and drawing capabilities. The drawing function in the REA version appeared to operate faster than in the original. Both versions also supported rotation and annotation saving.

Contextualizing REA: Not a New Breakthrough

A key point of contention is the perception of REA as a novel development. The underlying tools it utilizes are not new. Ghidra was released by the National Security Agency in 2019. Hopper Disassembler and IDA have been available for considerably longer periods, with IDA dating back to the 1990s. REA's primary function is to provide AI agents with a simplified interface to access these existing tools, thereby enabling reverse engineering.

The widespread enthusiasm for REA has been fueled by its viral spread on platforms like GitHub, evidenced by a rapid increase in its star count. However, critics argue that this focus on metrics like GitHub stars suggests a lack of compelling, substantive demonstrations of REA's unique value. The argument is that if the technology itself were sufficiently groundbreaking, the focus would be on its capabilities rather than its popularity.

While REA facilitates reverse engineering, the ability to reverse engineer software is not new. AI agents can be instructed to analyze specific features of an application and attempt to replicate them, even without specialized tools like REA. This can involve capturing animations or other visual elements through screenshotting and then using the AI's reasoning process to recreate them. While the accuracy of such AI-generated replications may vary, it is hypothesized that the resulting functions could be more efficient than human-developed counterparts. This was observed in the repurposing of the Photoshop suite, where vanilla AI models like Claude were able to achieve close approximations.

Broader Implications for Software Development

The trend towards AI-driven software reconstruction suggests a shift in the software landscape. It is posited that a significant portion of software will become open-source as development moves to a higher abstraction layer, focusing on general intelligence (e.g., Claude, Codex) capable of rebuilding software on demand. This could lead to the commoditization of front-end and back-end development.

From a business perspective, this trend presents both opportunities and challenges. While it lowers the barrier for recombining existing ideas into novel applications, it also potentially closes doors for certain types of software businesses. Wrapper applications that add functionality or visual enhancements to AI models may become obsolete as the underlying models are updated to incorporate these features directly. The value ladder for software development is perceived to be shifting.

While short-term benefits might be derived from reverse engineering existing applications, intellectual property restrictions remain a significant hurdle. The primary impact of this trend is seen as a reduction in the effort required to iterate on and combine existing concepts.

Furthermore, the necessity of specialized reverse engineering tools like Ghidra or IDA may diminish. Vanilla AI agents, through their inherent reasoning capabilities, are increasingly able to produce plausible visual and back-end analogs of applications. While these tools may offer speed advantages, it is anticipated that general AI agents will capture much of this functionality. The "reverse engineer era" is presented as an imminent development.