Agent delegates a large dependency audit to several subagents and reports back a summary. You then ask the same chat about a specific configuration file one subagent opened, and it has no knowledge of the file at all. What accounts for this?

The principle — Agent can spin up subagents to handle research, shell work or browser interaction in parallel. Each subagent runs in its own context window and returns a result to the main conversation. The separation is the point: it keeps a wide search from flooding the parent chat with everything it read.

Why the key is correct — Only the returned result crosses the boundary. Whatever the subagent opened, skimmed and discarded stayed inside its own window, so the main conversation can answer from the summary and nothing beneath it. That is why a detailed follow-up comes back empty even though the delegated task clearly succeeded.

Why the others are wrong — Context trimming does happen in long chats and would explain a gap, but nothing was trimmed here, because those messages were never in the parent window to begin with. A failed run would also explain silence, except the audit did report back; the summary arrived exactly as designed. And there is no filesystem barrier at all — the main agent can read that same file whenever you ask it to, which is precisely how you recover the detail.

Remember this — Ask a delegated task to return the specifics you will want later, because only its result survives the hand-back.

Sources — Cursor's Agent mode help documentation.