REA wants your AI coding agent to reverse-engineer anything - here's what that actually means

What’s actually been released

REA is a small, installable toolkit that hands your AI coding agent the ability to inspect compiled software, disassembly, decompiled pseudocode and call traces, and then describes what it finds in plain English. You install it with a single terminal command (npx rea-agents@latest setup), point your agent at a running program, and ask it questions. The project’s own site walks through two demo cases: working out why Windows Calculator gives you 220 when you type “200 + 10%”, and figuring out why the Chrome dinosaur game speeds up the longer you play.

Both are real, checkable examples of a genuine class of problem. Traditionally, answering “why does this program do that?” means manually stepping through x64 assembly, matching jump addresses to opcodes, and tracing function calls by hand, the kind of grinding, specialist work that keeps reverse engineers employed. REA’s pitch is that an agent can do the branch-decoding and call-tracing itself, and simply hand you the rule it recovered, in this case “take 10% of the first operand, then add it”.

The bit that’s still a claim

Here’s the catch: REA supplies the raw material, assembly listings, decompiled code, traced values, but it’s still the AI agent doing the interpreting. The worked examples on the site read cleanly because they’re curated showcases. Nothing in the source material demonstrates how reliably an agent reasons through messier, obfuscated, or deliberately anti-tampering code, which is most of the commercial software people actually want to understand. “Illustrative expert reasoning” is literally the label the site puts on its own walkthrough, a tidy reconstruction, not necessarily what happens on your first real attempt.

It’s also worth noting the discussion around this has clearly been lively: the Hacker News thread sits at 259 points and 83 comments, which tends to mean either genuine enthusiasm, genuine scepticism, or both. We haven’t seen anything in the source material that substantiates specific technical objections, so treat “this works brilliantly on real-world binaries” as unproven until you’ve tried it yourself.

So who does this actually affect

If you’re an ordinary reader who uses Calculator or plays the Chrome dinosaur game, nothing changes for you; these are demo targets, not products being broken. The people this matters to are developers, security researchers and hobbyist tinkerers who already do reverse engineering, or want to start, and who use AI coding agents day to day. For them, REA is a workflow shortcut, not a new capability out of nowhere: disassemblers, decompilers and debuggers already exist. What’s new is wiring an agent up to use them and narrate its findings, rather than you reading raw assembly yourself.

One point the site doesn’t address: reverse-engineering proprietary software, even for understanding rather than redistribution, sits in a legal grey area depending on jurisdiction and the software’s licence terms. That’s true with or without an AI agent involved, and nothing about REA changes those rules.

The takeaway

REA is a real, installable tool that automates a genuinely tedious task, getting an AI agent to read and explain compiled code, using two neat, verifiable demos. Whether it holds up on gnarlier, real-world software is untested by anything in the public material so far. If you’re the kind of person who’d reach for a disassembler anyway, it’s worth a look. If you’re not, this changes precisely nothing about how your calculator or your browser’s offline dinosaur behave.