What if the main coding-agent session was intentionally dumb?

Instead of making the main session the “brain”, what if it did almost nothing except call a subagent?

The subagent checks the project:

What’s already done?
What’s broken?
What’s next?
What should I work on?

Then it does the work, updates the project state, and finishes.

The main session doesn’t need to understand the project.

It doesn’t need to maintain a huge context.

It doesn’t even need to meaningfully process the subagent’s response.

It can basically just keep calling:

“Go check what needs to happen next.”

Main session = runtime.

Subagent = brain + worker.

Project state = memory.

I’m wondering if this could make long-running coding agents much simpler and avoid massive persistent contexts.

Thinking about this for Claude Code, OpenCode, Codex, Hermes Agent, Cursor, Cline, Roo Code, OpenHands, Aider, Goose, Gemini CLI, and similar agents.

Has anyone built this pattern already?