You ask the agent to find every place user input is validated. It spawns an Explore subagent, which reports back, and the agent then proposes an edit to one validator file. Before accepting, you want to know whether the main agent has actually seen that file's contents. What did the subagent hand back?

The principle — The Explore subagent runs in its own context window and reports back. It keeps the main conversation focused by summarising results instead of dumping raw file contents. That is what makes it cheap in context, and it is also what makes its output lossy by design.

Why the key is correct — What reached the main agent is a set of findings, not the files behind them. Anything the subagent judged irrelevant — a guard clause, an early return, a second call site in the same file — is simply absent from the main context, and the proposed edit was written without it. When the exact contents of one file decide whether a change is safe, open that file or attach it to the main conversation so the model doing the editing has genuinely read it.

Why the others are wrong — Assuming a verbatim copy inverts the design; if raw contents flowed back, spawning a subagent would buy nothing. Assuming a shared context window makes the same mistake from the other direction, since the separate window is what allows many parallel searches without bloating the thread. And assuming the findings are thrown away goes too far the other way: they are returned and acted on, which is precisely why a summary-based edit can look well-informed and still miss a line.

Remember this — Treat a subagent's report as a map, not the territory. It tells you where to look; it is not evidence that the main agent has read what is there.

Sources — Cursor's Search tool documentation.