Peeking into AI coding agents
I’ve had a bit of free time in the past few days and I’ve been reading about AI coding agents.
It looks like in the end it’s mostly sophisticated wrappers around LLMs, with the main advantage being access to tools they can use to gather more information or execute actions.
These are all just plain text: you tell the LLM how to communicate if it wants to use a tool, then when requested you run them and feed back their output.
I guess people tried this already in the past, but it got traction only now as the major providers’ models (Claude, GPT, Gemini) got really good at using them.
I’ve been playing around with aider and cursor and, as of today, the latter seems to overall be more “intelligent” only because it has access to a wider set of tools (most notably text search).
It’s also interesting to look at the prompts these are passing to the LLMs.
Aider’s ones can be found in its
repo (that’s the
nice part of open source, one can learn so much by looking at existing
projects), cursor however hides them… unless you use
gemini-2.5-pro-exp-03-25 and ask it to return the entire content of the
request you’re sending (here’s the chat
log).
That seems to be a vulnerability of some sort in this model, I couldn’t get Claude or GPT-4o to do the same. Or maybe that’s just an hallucination, who knows.
Anyway, I hope in the future we’ll move away from this trend of hiding system prompts, and it will become normal to expose these and allow their customization by users.
Further reading:
