October was the month of "I reverse-engineered X with AI." GTA V running on an iPhone, open-source clones of Adobe's suite, ports of old games. Most of those headlines contained the letters REA.
On November 10 the maintainer of Awesome REA went through 39 of them one at a time. The method was unglamorous but sound: read the project's own README, its agent instruction files, the author's post and REA's official showcase, and look for a primary source that says REA was used.
The tally: 26 confirmed, 3 partial, 4 unverifiable, 6 that were not REA at all.
The most awkward part is the top of the list. Of the six cases with the largest reach, only DX-Ball states in its own repository that it used REA. Two more mention it only in the author's post. The remaining three, including a modding tool with about 14,000 likes and a native Plex client for LG televisions, used other tools or could not be sourced at all.
In other words, a good share of what circulated as "REA's greatest hits" was Claude Code or Codex driving Ghidra and IDA directly, or plain manual reverse engineering.
Two controlled comparisons, neither flattering
The first is rea-lab: five preregistered studies, 482 runs, Claude Code throughout. Not one study found a significant improvement in solve rate from using REA. The REA runs took roughly twice as long and cost more.
The second is leanrea: 36 runs against REA 6.5.0. A plain CLI skill scored 74 out of 75, against REA's 75 out of 75, while using 38 percent less wall time and 35 percent less money.
Both are small, single-model studies. Treat them as hints. But they point the same way: the agent and the underlying engine do most of the work, and the orchestration layer contributes less than the star count implies.
An independent hands-on test said much the same
On October 12 someone ran REA on Debian 13 with 8 vCPUs, no GPU and no cloud API keys. The target was a 100MB Electron app that unpacked to 459 files, plus a five-file toy app with known answers. Static JavaScript analysis came out fast and solid, but missed two preload files that were plainly visible with grep. The other finding was call burden: a single investigation issues dozens of tool calls.
For a harder datum, SRE-Bench reports that the best agent fully solves 31.5 percent of realistic test binaries. That is the ceiling with the strongest agent and the strongest tools, not the floor for anyone who installs a package.
And there is a hidden bill: context
Someone measured the tool surface at version 6.7.0: 207 tools, roughly 179,000 tokens by cl100k estimate, about 21 times an 8k context window.
That is not a defect. It is what happens when you hand an agent an entire professional toolkit. The project noticed: since 6.8.0 the MCP input schema defaults to a compact profile specifically to reduce that overhead.
So should you install it
Yes, but put it in the right place. What it removes is the dullest part of the work: exporting pseudocode by hand, pasting it, losing the thread, starting over. That saving is real.
What it does not replace is judgement. Which function is worth following, what a branch implies, whether the recovered logic can be rebuilt as your own implementation. The two comparisons say the same thing: orchestration is not a moat, understanding is.
The next piece is operational. It starts with the reason most installs fail.
Of 39 Viral "Built With REA" Demos, 26 Actually Used It
← Previous: REA Went From 10,000 Stars to 451,000 in Five Weeks by Wiring Ghidra and IDA Into Your Agent
Next: →Getting REA Running: Clear the Node Version Wall First →