Many of you who code your own QuickApps (or modify existing open-source ones) have probably realized by now that a coding agent can help a lot — GitHub Copilot, Claude, ChatGPT, DeepSeek, or whichever AI coding service you use.
When I first tried it, about two years ago, it was hit and miss. Sometimes the code was fine, sometimes it was not. I saw particular issues with Lua — probably because Lua is not the biggest language on GitHub, and the agents were mostly trained on the more popular languages like JavaScript and Python.
By now, in 2026, the Lua code agents produce is quite good. As an old university teacher of programming methodology, I do not have much to complain about. What is missing is two things: the domain knowledge to code QuickApps for the Fibaro system, and a fast code → verify loop to arrive at a correct solution.
Some of it has crept into today's agents already: even without any
guidance they know roughly how a QA should be coded, so they must have
trained on the QA/Lua code that is out there. But with some added
SKILL.md documents outlining specific techniques, issues and
gotchas for QuickApps — together with AGENTS.md directives for
QA development — the output from the agent can be very good:
"Please code a Philips Hue QA for me — the credentials are in my .env file, and here is a link to the new V3 Hue protocol. Outline a plan for how to do it, and create a minimal viable QA to test it."
A prompt like this gets you very far — if the agent has been given the right skills and tools.
An excellent tool for the agent is my flua QA emulator. flua is a Lua engine — a bundled Lua interpreter with a command-line interface — with a built-in emulation of an HC3. Feed it a QA file and the QA runs offline:
flua --api local --instant hello_hue.lua
This runs flua against the simulated HC3, fires all timers faster than
real time (while keeping a virtual clock), and emulates the QA
hello_hue.lua. Exactly what the agent needs: the
coding→testing cycle should not involve the real HC3 — your family will
not appreciate the lights going wild while the agent iterates.
And the agent does not just write the code — it verifies it, in a loop, against the emulator, without a controller anywhere in the loop. That is what makes the partnership fast.
AI coding is here to stay, it is getting increasingly good at making
QuickApps — and with tools like flua the process is becoming even more
agent-friendly and streamlined. Give your agent the skills, an
AGENTS.md (flua --tool setup does just that), and flua, and you have a coding partner that both
writes and verifies the code.
flua: https://github.com/jangabrielsson/flua — QuickApp emulator for offline development. I am also posting a tutorial series here on the forum, from your first QA to agent-driven development.