Now that you understand how your agent behaves under normal conditions, let's explore some practices that can improve its output.
I want to talk about plan mode - a feature I used to recommend enthusiastically, but no longer recommend in the same way.
Most agents now ship with plan mode. Claude Code, Codex, and others all have it. You can build your own version in other tools as well.
The point of plan mode is simple: create a plan before implementing anything.
You saw this in the previous video when the agent automatically entered plan mode after detecting a planning conversation.
In the default implement mode, your agent can do four things:
Plan mode removes the ability to write files. The agent can still read files, run bash scripts, and call MCP servers - but it can't make any changes to your codebase.
The concept of planning is powerful.
Talking through what you're building helps you understand it better. It also aligns you with the agent.
A deliberate planning step forces exploration. The agent explores the repo, gathers context, and figures out what it needs to implement properly.