People are using coding agents to reverse-engineer home devices and write automation for them in a few hours. Work that previously required days of analysis on undocumented protocols, and that almost nobody tackled because it wasn’t worth the effort.

Simon Willison collects the anecdotes and draws out the real point: the cost of writing code has collapsed, and this changes the equation.

Before, understanding how an undocumented device communicates was feasible but uneconomical. The time spent deciphering the API exceeded the value of the automation you’d get out of it. There was a second problem too: undocumented APIs change or break, and all your work had to be redone.

With coding agents the calculation shifts. The cost of a first attempt is low, and the cost of trying again if it fails is equally low. If the API changes in six months, throwing away the code and rewriting it costs a fraction of what it used to.

Why this matters to you. If you own a device that does something useful but the maker’s app is terrible or locked down, the threshold for when it’s worth writing your own automation has dropped. You don’t need to be a reverse-engineer: you capture the device’s network traffic (tools like Wireshark or an HTTPS proxy) and tell the agent what you want to achieve.

The caveat is fair: these are anecdotes, not benchmarks. We don’t know how many devices lend themselves to this treatment or how well the code holds up over time. But the signal aligns with what we covered on July 18th: coding agents have moved past the demo phase and are changing what’s worth doing.

In detail

How it used to work.

Reverse-engineering a home device has always been possible. Turn on Wireshark, capture traffic between the maker’s app and the device, study the message patterns, write a script that reproduces those messages. It works. People were doing this ten years ago.

The problem was ROI. Deciphering an undocumented protocol takes hours, sometimes days. And any experienced programmer knows that undocumented APIs are unstable: the maker can change them with a firmware update, and then all your work needs redoing. You’re locked into a frustrating maintenance cycle just to control a lightbulb.

What’s changing.

Coding agents lower two costs at once. The first is the cost of the attempt: instead of spending an afternoon reading binary logs, you describe what you captured to the agent and ask it to write code that reproduces the behavior. The second is the cost of failure: if the code doesn’t work, you throw it away and try again, because rewriting costs little.

Willison’s point is psychological as much as economic. When writing code costs little, the prospect of maintaining it later (or ditching it and starting over) stops weighing you down. The mental friction that held back DIY dissolves.

How it works for non-developers.

A smart device communicates with its app by sending messages over your home WiFi. If you intercept those messages, you see a pattern: send these bytes and the light turns on; send those and it changes color. A coding agent can read a log of that traffic and write a script that does the same thing, bypassing the official app. The result is your own automation, which you can trigger with a voice command, a timer, or a system script.

What we don’t yet know.

The source is a blog post collecting anecdotes, not a systematic study. Three practical questions remain open.

First: how many devices are suitable for this treatment? Many makers now encrypt traffic between app and device, and if the protocol is encrypted, interception becomes much harder.

Second: how well does the code hold up over time? A firmware update can change the protocol, and your automation breaks. Willison’s argument is that rewriting it costs little, but it remains to be seen how often this happens in practice.

Third: nobody has yet measured how long it actually takes. The anecdotes mention “a few hours,” but there’s no large enough sample to know if that’s the norm or the lucky case.

What is certain is the direction: the threshold for what’s worth doing has dropped, and it will drop further. If you own a device you’d like to control better, it’s worth trying.