Research - Plan - Implement

I found this methodology fascinating:

The Three Phases

The workflow consists of three phases, each in a separate session/chat:

Research: Search the codebase, gather context about the problem, document everything in detail
Plan: Generate an implementation plan
Implement: Write the actual code

Context Management Is the Core Idea

The underlying concept: Every model has a limited context window. The more of it you use, the worse the results—the rule of thumb is that quality degrades above 50% context utilization. By splitting work into separate chats, each phase stays “fresh.”

Where the Leverage Actually Lies

What surprised me: The highest leverage is in the research phase. Then comes the plan. The code itself? They sometimes don’t even look at it anymore.

The review process reflects this: A colleague reviews research and plan — but usually not the code.

Sub-Agents and Their Advantage

The big advantage of sub-agents: They have their own context. You give them a task, they complete it, and return only the result. All the tool calls they make along the way don’t end up in the main agent’s context.

Resources: