Digital archaeology usually involves painstaking manual labor, where developers spend weeks reverse-engineering undocumented binaries and guessing the intent of programmers who retired decades ago. The process is often a gamble, balancing the risk of breaking legacy dependencies against the need for modernization. This week, however, the boundary between manual recovery and automated migration shifted significantly as a 33-year-old piece of software was resurrected not by a human expert in obsolete hardware, but by an AI agent.
The Binary-Level Reconstruction of Babylonian Twins
The project centered on Babylonian Twins, a game originally developed in 1993 for the Amiga 500. The task was to migrate this relic into the modern Godot 4 engine using Claude Fable 5 operating within the Claude Code environment. The scale of the challenge was immense: the model had to ingest and analyze 72,758 lines of 68000 Assembly code, the vast majority of which contained no comments or documentation to guide the process.
Rather than simply attempting to mimic the game's behavior, the workflow implemented a rigorous verification loop to ensure absolute fidelity. The AI utilized vasm, a modern assembler, to rebuild the 1993 source code. It then employed diff tools to compare the resulting output against the original released binary, ensuring a byte-for-byte match. When the model encountered syntax discrepancies between different assemblers, it did not halt; instead, it autonomously designed and applied a preprocessing stage to bridge the grammatical gap between the old and new tools.
This technical precision extended to data recovery from compressed binary maps that lacked headers or dimension data. Claude Fable 5 reverse-engineered the tile grid structures, physics properties—such as collision boundaries and climbable surfaces—and the specific behavioral patterns of enemy characters. It even identified and recreated hardware-specific effects known as Copper lists, which the Amiga used to generate sky gradients and water color cycles. The final result is a unified application that allows users to toggle between the original 1993 version and the 2026 Godot port in real-time.
From Pattern Recall to Agentic Reasoning
The significance of this port lies in the transition from recall to reasoning. Most LLMs rely on patterns found in their training data, but the volume of Amiga 68000 Assembly in modern datasets is negligible. The fact that Claude Fable 5 succeeded suggests it was not remembering a solution, but logically deducing the operational principles of the hardware and code in real-time. This marks a departure from AI as a knowledge base and a move toward AI as a logical engine capable of analyzing unfamiliar systems.
This shift fundamentally alters the developer's role. In this workflow, the human ceased to be the primary coder and instead became the Decision Maker. The developer provided high-level qualitative feedback, noting that a jump felt off or a collision felt unnatural. The AI then analyzed these subjective complaints to find objective technical causes, such as the friction calculation differences between 60Hz and 50Hz tick rates, and implemented the necessary mathematical corrections.
Beyond the code, the automation extended into the operational pipeline. The AI handled the tedious tasks of the release cycle, including executing the game to move characters to specific coordinates and capturing screenshots for store pages in 11 different languages. Because the Steam dashboard lacks a public API for certain metadata, the AI directly controlled the browser to input data and upload files. This demonstrates an evolution where the AI manages the entire lifecycle—planning, development, QA, and deployment—rather than just writing isolated functions.
While the results are impressive, the process revealed the remaining boundaries of AI autonomy. The model struggled with logical edge cases, such as incorrectly setting the push range for guard characters, which allowed them to attack players through walls, or miscalculating sound loop lengths, leading to audible gaps. These errors confirm that while AI can handle the bulk of the migration, domain-specific human testing remains the only way to resolve nuanced logical failures.
For engineers dealing with undocumented legacy systems, the lesson is clear: the metric for AI utility is no longer whether it can write code, but whether it can reason through a complex legacy context and build its own verification environment. By granting the AI terminal access and file system permissions to build, run, and test its own hypotheses, the development process transforms into an agentic workflow that drastically reduces the risk of system modernization.
