This weekend, developer Ammaar Reshi gave OpenAI's Codex 5.4 a compiled DOS binary and said: figure it out.
No source code. No documentation. Just the executable.
Six hours later, the agent had unpacked assets, disassembled the EXE, reconstructed the renderer logic, and rebuilt SkyRoads — a beloved 1993 DOS game — from scratch in Rust. The whole thing ran while Ammaar was doing something else. He posted a video. People couldn't look away.
Neither could I.
What Actually Happened Here
This wasn't a party trick. Walking through what Codex had to do makes the result more impressive, not less:
- Unpack compressed assets from a 30-year-old binary format
- Disassemble x86 machine code with no debug symbols
- Infer game logic from raw CPU instructions
- Reconstruct the rendering pipeline from memory addresses and draw calls
- Re-implement the whole thing in a modern language (Rust) that didn't exist when the original was written
Humans do this kind of work — it's called reverse engineering, and it takes experienced engineers days or weeks. Codex did it in six hours, unattended.
Why This Matters Beyond Nostalgia
Legacy software is not a retro gaming problem. It is one of the most common pain points I hear from small business owners.
The machine shop running scheduling software from 2004 that nobody touches because the guy who wrote it retired. The regional distributor with a custom inventory system built on Access 97. The law firm with a billing tool that only runs on a Windows XP virtual machine because that's the only thing that can open the database. The manufacturing company that depends on a custom ERP that nobody has the source code for anymore.
Every one of these businesses is one hardware failure or OS upgrade away from a serious operational crisis.
Until recently, the options were grim: pay a reverse engineer $200/hour to document the thing, rip and replace with an off-the-shelf product that doesn't quite fit, or keep running the ancient system and hope nothing breaks.
What Changes With Long-Running Agentic Coding
The Codex demonstration points to something that will matter more in the next 12-24 months: AI agents that can work on hard, ambiguous problems for extended periods without supervision.
That's different from "AI that writes code." Writing code from a clear spec is useful. Working autonomously through undocumented systems — reading what's there, forming hypotheses, testing them, iterating — is a different capability class.
For small businesses with legacy software, this starts to open up options that weren't cost-effective before:
Documentation extraction. Even if you don't rebuild the software, having an AI agent read a legacy binary or database and produce human-readable documentation of what it actually does is valuable. You'd know what you have before something breaks.
Targeted modernization. Rather than a full rewrite, an agent can identify specific components — the report generator, the import/export module — and rebuild just those in something maintainable, leaving the rest intact.
Migration scaffolding. Moving from an old custom system to a modern platform (QuickBooks, NetSuite, anything else) requires understanding what data you have and how it's structured. An AI agent can work through undocumented schemas and produce migration maps.
None of this is fully automated yet, and the Codex demo was a controlled showcase. But the underlying capability — sustained, autonomous reasoning about complex undocumented systems — is real and improving.
What to Do With This Today
If you have legacy software you're worried about, the most useful thing you can do right now is create a risk register. What systems are you dependent on? Who built them? Do you have the source code? What happens if it stops working next Tuesday?
That inventory doesn't require any AI tools. But once you have it, you're positioned to have a much more specific conversation about where agentic tools might actually help — and where a human engineer still needs to be in the loop.
The six-hour DOS session was striking because of how autonomous it was. But the more important shift is that these long-running, multi-step, ambiguous tasks are now within reach for organizations that couldn't afford weeks of specialized engineering. That changes the calculus on legacy modernization in a way that matters for small businesses specifically.
Curious whether AI-assisted modernization makes sense for a legacy system you're running? We work through exactly these questions with small business clients. Reach out and we'll take a look.