Google’s terminal agent competes on a specific axis: how much of your project it can consider at once.
When that genuinely helps
Some problems are spread out. “Why does this value end up wrong on the settings page” can involve a form, a store, an API route and a database helper — four files that only make sense together. If the agent can hold all four plus their surroundings, it can reason about the actual path instead of guessing at the edges.
Large context also helps with the unglamorous work: understanding an unfamiliar codebase, finding every place a pattern is used, and migrations that touch dozens of files in the same way.
When it does not matter at all
Most day-to-day tasks are small and local. Fix this function. Add this field. Rename this thing. For those, a huge is capacity you are not using.
Quick check
You are adding one field to a form and updating the code that saves it. Does a huge context window help?
The trap: capacity is not attention
It is tempting to think “more context = better answers”, so you paste in everything. In practice, a long undifferentiated dump can make results worse: the important detail is now one line among thousands, and the model has more chances to latch onto something irrelevant.
The skill — with every agent, not just this one — is choosing what goes in. Point at the three files that matter. Say what the constraint is. Name the thing you already tried.
A rule that holds across every tool
Curated context beats maximum context. If you find yourself hoping the window is big enough, that is usually a sign the request should be smaller.
Next: the editor-shaped option, and where Grok and Kimi fit into all this.