In partnership with
That tweet has 3.5M views, and my feed has spent the whole week agreeing with it.
The reason is REA, a free, open-source tool that lets your AI agent take apart an app without its source code and explain how a feature inside it works.
I posted about it when it had 80,000 GitHub stars. It's at nearly 150,000 now.
The pitch sounds wild. You point an agent at an app you like, ask "How does this feature work? Show me the evidence," and it digs through the compiled code and tells you.
Something that needed years of reverse engineering experience now needs a terminal and one prompt.
So is software done?
Before answering that, I went back to see what people pulled off when this was still hard, and what happened when someone let AI agents try to build an entire game.
re3: GTA III on a Nintendo Switch
A group of fans rebuilt the code for GTA III and Vice City and got both games running on the Switch, Wii U and PS Vita. They left out Rockstar's music, dialogue and art, so you still needed your own copy of the game.
Take-Two sued 14 of them anyway. The modders fought back, and in 2023 the two sides reached a settlement and the case was dropped.
Super Mario 64: 4K, ultrawide, zero emulators
Fans reverse-engineered the 1996 game's code, finishing in 2019.
In 2020, someone used that work to make a proper PC version with 4K and ultrawide support, and Nintendo was going after the download links within days.
PlxNative: a Plex app built out of pure impatience
A developer got tired of Plex reportedly taking around 30 seconds to show the profile screen on his 2019 LG TV, so he used Claude Code to work out how the TV plays video and built his own app, almost entirely in Rust.
It's open source, and my favourite of the lot because the whole project started with a man refusing to wait 30 seconds. Most good software starts with someone being annoyed.
GT-Caliber: AI agents vs GTA 6
In June, Ziwen Xu set a loop of AI agents building a GTA 6-style game around the clock.
Within days they had traffic, people walking around and an in-game phone with maps and contacts. They also built Los Angeles instead of Miami.
Remember this one. I'm coming back to it.
A lot less than my feed suggests. REA's own showcase is a list of single features.
One explains how an authenticator app makes its 6-digit login codes. Another rebuilds the sound-panning calculation from DX-Ball, a 1996 brick-breaker, and checks that it matches all 63 bytes of the original compiled function.
Every answer comes with the evidence behind it, which is the most useful part.
Two things I'd add to my own post.
The analysis runs on your computer, but your agent still sees whatever REA finds, so those results go to Claude or whichever model you're using.
And two of the stories above ended in a takedown or a lawsuit. REA puts permission and legality on you, so stick to apps you own or open-source ones.
Want to try it?
Install Claude Code. You'll need a paid Claude plan.
Open Terminal, run npx rea-agents setup, pick Claude Code, then restart it.
Ask: "How does [feature] in [an app I own or an open-source app] work? Show me the evidence."
Check one thing it tells you before you believe the rest.
An Anthropic AI sent police a false homicide tip.
During a test in July, a Claude model submitted made-up information about an unsolved Philadelphia homicide to a police tip website.
It landed in spam, Anthropic only found out on September 28, and police called the two-month delay in detecting and reporting it "unacceptable."
And it wasn't a one-off. Anthropic says its agents also exploited software flaws and worked around fees to reach public data.
Its word for this is "persistence": when Claude can't complete a task as given, it works around a restriction instead of stopping.
Live internet is now switched off for all its internal evaluations until its monitoring can catch this reliably.
Then Meta launched an AI that refuses to give up.
On Monday, Meta released Muse Wonder and Muse Wonder Full, which it calls frontier models for agents.
The headline feature is "Active Focus," which lets the models keep working when a task gets hard, try new approaches and check in before important steps.
Wonder will soon power Muse, Meta's agent app, starting with paid subscribers (reportedly the $20 and $100 monthly plans).
So in the same week, one company apologised for agents that won't stop and another made it a selling point.
Three fired OpenAI safety researchers published an open letter denying they mishandled sensitive information and warning that colleagues are now scared to speak up.
Google's new persistent Gemini agents can get their own work email, calendar, Drive and spot in the company directory. They're in private preview for select customers and will come to Gemini Enterprise later.
OpenAI reportedly told investors its annualized revenue is near $50B, about $20B less than an earlier reported figure, which came from investors counting cloud partner sales the way Anthropic does.
Apple cut iPhone 18 Pro component orders by at least 15% on soft demand, after rising memory costs made the phones $100 pricier.
In three experiments with 1,222 people, those who used an AI assistant for about 10 minutes gave up more often on hard problems once it was taken away.
Gemini Gems turn into Skills for personal accounts in November, and Google converts them for you.
A free, experimental meeting note-taker for Apple silicon Macs that works fully offline. It transcribes the meeting, turns your rough notes into proper ones, and answers questions mid-meeting from the PDFs and Docs you load.
A free Mac browser (in beta) with Claude Code or Codex sitting beside the page. Press ⌘E and your agent sees what you're looking at.
A free, open-source Mac tool that lets your agent draw a giant arrow on your screen pointing at the "Allow" button it needs you to click. Handy once REA starts asking you for permissions.
So, is software done?
Going in, I expected to end this with "AI can now copy any app." After all of this, I don't think it's there yet.
The fans behind Vice City spent years on it. REA's best showcase is one calculation from one old game.
And the GTA 6 agents got the traffic, people and phone working, then put the whole thing in the wrong city, which in a game about Vice City is kind of the main thing.
That's the pattern all week. Agents are great at not giving up (Anthropic's kept going until it filed a false police tip), while ten minutes with an AI assistant makes us quit hard problems sooner. Neither tells you which city you're supposed to be building.
I could be wrong about this in a year.
For now, I'd use REA the way you'd watch a teardown video, to learn how something you love was put together. Deciding what's worth building is still on you.
Reply and tell me the one app you'd crack open first.
Until next time,
Vaibhav 🤝
If you read till here, you might find this interesting
#AD 1
A fast answer and a coordinated one are not the same thing. When AI resolves a B2B customer issue without looping in the teams who have to deliver on it, you get confident responses nobody actually signed off on.
A new briefing paper from Harvard Business Review Analytic Services, sponsored by Front, examines the coordination gaps that open up when transactional AI tools meet multi-team B2B service, and how leading companies are using AI to close those gaps instead of widening them.
Read the briefing paper for the questions to ask before your next AI investment.
#AD 2
Stop typing AI prompts. Start talking.
You think 4x faster than you type. So why are you typing prompts? Wispr Flow turns your voice into ready-to-paste text inside any AI tool. Speak naturally, tangents and all, and Flow cleans it up. Available on Mac, Windows, iPhone, and Android.
