I’ve noticed a shift in how people are using coding tools recently. Even developers who used to rely heavily on full IDEs seem to be moving toward lighter editors, terminal workflows, or chat-based environments, with agents doing more of the actual work.
That direction makes sense in a way. If the agent is doing most of the editing, navigating, and fixing, the human no longer needs as much of the IDE’s support. The value of inspections, refactorings, and deep project awareness becomes less visible if you are not the one directly interacting with them.
But I’m not sure that means those capabilities have become less important. It may just mean they are no longer being used.
A full IDE like IntelliJ still maintains a much richer understanding of the codebase than what most agents currently operate on. Types, structure, references, warnings, refactorings, test runs, and project configuration are all already there, continuously updated.
The question, at least to me, is whether agents should be working through that model instead of approximating it through text edits and incremental discovery.
My current intuition is that agents might actually benefit from these tools at least as much as human developers, possibly more. A deterministic refactoring engine, immediate inspection feedback, and access to the project structure seem like things an agent could use very effectively, especially compared to repeatedly rediscovering the same information from raw files.
This is the direction I’ve been pushing with AgentBridge. It is a working plugin at this point, with a couple of thousand downloads, and builds on top of what JetBrains already exposes while trying to shorten the feedback loop even further and keep the interaction closer to how humans already benefit from the IDE.
It is still early for agentic coding in general, and I don’t expect a single approach to win. Lightweight tools, external agents, and IDE integrations will likely coexist. This is simply one direction that seems worth exploring more deliberately, especially in environments where the IDE already understands a large part of the codebase.