Agents don't feel like chatting

As of today, the main way we interact with agents is via a chat interface. They are also triggered in automated workflows, but the most common way is in chat. In an usual development loop, I have some kind of spec of document deliniating what should be implemented, and I feed it to the agent. Yeah, sure, or have an agentic workflow that spawns another agent and reviews, or multiple agents that reviews through a specific lens, you can go pretty crazy here. The point is, at some point in time, you will stop burning tokens and, albeit a bit controversial, you are going to review the work.

Maybe I am holding agents wrong, but I very frequently see issues in the implementation, and I want to get into a quick loop with the agent where I ask a question, it answers, potentially fixes it.

It doesn't really happen that way. Say, I notice that the implementation added some log output that's out of place, or it's in a hot path and it will become very noisy. I tell that to the agent and I get a wall of text. Sometimes, inscrutable, using terms that weren't present in the chat to refer to abstractions in code.

It feels like chatting with a co-worker who is trying really hard to sound smarter than everyone else. Maybe the AI labs are being incentivized to make their agents behave that way: they do want you to believe that the agents are smarter than you.