- A hobbyist's account says an AI agent built a Kindle photo app that passed automated checks but left traces of old images on the physical display.
- The agent made progress once it could read logs and capture the screen itself. That feedback loop, not the code, was the turning point.
- Ask of any AI-built system: what does the agent check against reality, and who defines 'done'?
Software that passes its tests is not yet finished
Most talk about AI coding agents is about how fast they write code. A small experiment suggests the harder part is knowing whether the code worked. The lesson here: an agent is only as good as what it can observe.
Niraj Rajgor, on his own blog, describes turning an unused 10th-generation Kindle Basic into a photo frame with coding agents on his Mac. He called the project KindleCTL. This is one person's account of one device. It is not a study.
What the experiment built
Rajgor started small. He wanted his own image as the Kindle's screensaver. Then he asked for a slideshow app. It now opens from the same screen where he picks books. It cycles through ten photos, showing each for four minutes.
To run his own software, he needed to jailbreak the device. That means removing the maker's limits so custom programs can run. He says the work used existing open-source jailbreak code from SpiderCat, released by sparklerfish and the KindleModding community, as a reference. The setup happened entirely over Wi-Fi, with no USB connection.
Three tools, three different outcomes
Rajgor reports that his first agent, GPT-5.6 Sol, loaded a helper onto the Kindle. But it could not find a place on the device to store content. He then tried Fable 5.1 and Opus 5.5. He says Anthropic's cybersecurity safeguards blocked each attempt after two prompts.
GPT-6 Astra, which he tried last, completed the jailbreak. Rajgor says he told it the software was for personal use and asked it to draw on open-source projects.
One user's experience on one device does not rank these tools. It does show something for security leaders. The same request can be blocked by one product's guardrails and completed by another. The controls that decide this sit with the vendor, not with you.
The feedback loop was the turning point
Early on, the agent could connect to the Kindle but could not display data. To debug, it asked Rajgor to photograph the screen with his phone. He found this tedious. He asked the agent to set up a way to read logs itself.
The agent built a persistent diagnostic connection. It could collect errors and test fixes without Rajgor downloading and reopening a package each time. Later, it added remote screen captures. Rajgor calls this the turning point that let the agent keep working on autopilot.
He also notes the agent had not suggested this on its own. It built the loop once he asked.
Where the tests fell short
The slideshow passed its automated checks on the Mac. On the Kindle, switching photos left traces of the previous image. The tests had not caught how the physical display refreshed. The screen captures showed complete images, but they could not show the traces on the glass.
The fix was to use the Kindle's built-in refresh mechanism to clear each slide. Rajgor then confirmed by looking at the device that each photo was clean. Rajgor calls the move from passing tests to something usable the most interesting part of the experiment.
This is a pattern leaders will recognise from any system. A dashboard can be green while the customer sees something else. An agent can only verify what its sensors can see. Here the sensor missed a flaw that a person caught by looking.
Questions to ask your teams
If your teams use coding agents, ask what each agent can observe about the real result. Can it read logs and see the live output, or does a person relay what happened?
Ask who confirms the final result against reality, and how. Automated checks are a start. They are not the same as a person checking the outcome.
Ask how your policy handles dual-use requests. Rajgor's disclaimer says to experiment only with devices you own or have explicit permission to modify. Your staff need the same rule, written down.
Rajgor's own advice for next time is to set up the feedback loop from the start. That is good advice for any AI project: give the agent eyes first, then ask for the work.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: nirajrajgor.com.
