AI coding agents are crude, but also beautifully simple. I’ve noticed a couple of things when working with both Antigravity and Claude Code. Their approach to debugging, or figuring out how to implement something involves dumping things into the context. This is why you often see coding agents adding fmt.Println() statements everywhere, running the code, and then subsequently removing them. This is how the agent “sees” what is going on.

I noticed this from Antigravity earlier:

~/Code/things-mcp $ grep -r "ServeSSE" vendor/github.com/mark3labs/mcp-go/server

Instead of rich insight from the toolchain using the language server protocol and instead of rich information from things like debuggers, the agents resort to grepping for strings in the output. In this case it is searching a directory for a mention of a specific string to see if the package implements functionality relating to serving SSE.

I should have expected this, it’s often what I do. But by the same token, the hype would have you believe that these agents are smarter than we are.