Alleged spyware-like hidden backdoor in Claude Code: analysis claims it checks timezone/region/proxy signals and injects them into system prompts through heavily obfuscated binary logic paths.
Every programmer should write their own agent.
It's fun, only takes 50 lines of code, and will surprise you. But to get the most from the experience, do these two things.
First, hand-code it from scratch. Start with a blank text file and type out every line of code yourself; don't use any AI, not even autocomplete.
Second, rely only on your language's standard library docs and the docs for Anthropic's API (or whatever model you're using).
You can easily prompt Claude to do this for you. But the value of building your own agent by hand is that it helps you build a strong mental model of how they work, not just in broad strokes ("they're just an LLM in a loop with tool calls") but in the nitty gritty details. By doing everything manually you can make sure you don't accidentally gloss over important bits you don't understand yet, and you can experience directly how quickly new behavior emerges.
For instance, my first agent had a single tool for adding two numbers. But that was enough for it to do multiplication by repeatedly invoking the tool.
Once you get something working, there are dozens of fun things to try: add a second tool (perhaps `bash`? Make sure to do this in a sandbox), swap out the model (what's the worst model you can use and still get good results?), or make a simple memory system.