Two developers chase the same bug. One asks the agent to look at whatever is wrong with the payment retry logic; the other writes that PaymentFailedError is thrown twice on retry inside PaymentService. The second run lands on the offending lines in a fraction of the steps. What accounts for the gap?

Hardcursorsearchgrepprompting

Last reviewed 21 Aug 2026

Referencing specific symbols makes Agent use grep automatically, turning the hunt into an exact match on a name instead of a broad semantic sweep over the whole codebase.Correct answer
Grep runs only when a developer invokes a search by hand, so the gain came from the second prompt being shorter and easier for the model to parse before it started reading files.
The second prompt hit a cached result left by the first developer's run, because search results for a repository are shared between workspaces and sessions once the codebase has been indexed once.
Search in Cursor is embedding-based whatever you type, so both prompts ran the same semantic retrieval and the second developer simply drew a luckier ordering of the results that came back.

Show answer & explanation

Correct answer: A. Referencing specific symbols makes Agent use grep automatically, turning the hunt into an exact match on a name instead of a broad semantic sweep over the whole codebase.

Explanation

The principle — Cursor does not have one search. The fastest way to find code is an exact match on a function name, variable, error string or regex, and Agent uses grep automatically when you reference specific symbols. Which search it reaches for is decided largely by what your prompt gives it to match on.

Why the key is correct — The second prompt contains two literal identifiers. That is enough for the agent to construct a pattern and go straight to every occurrence, including references across files, rather than describing the problem to an embedding index and reading whatever comes back. Cursor's Instant Grep supports full regex and word-boundary matching, so patterns like an import line combined with a class name can trace usage precisely. The first prompt offers no literal to match, so the agent has nothing to grep for and falls back to semantic exploration, which is the slow path when the answer is a known name.

Why the others are wrong — Believing every search is embedding-based makes the difference look like chance, and that belief is expensive: it removes the one thing you control. Believing grep is a manual tool you must type yourself misses that the agent selects it on its own, and it also fails on the evidence, since the faster prompt is the longer one. And attributing the speed to a shared result cache misreads what team index sharing does; a second developer's run is a fresh search, not a replay of the first.

Remember this — Before you describe a symptom, look for a name. An error class, a function, a config key, a literal string from the logs — one of those in the prompt turns a search into a lookup. Save the prose description for when you genuinely do not know what the thing is called.

Sources — Cursor's Search tool documentation.

Sources

Primary sourceSearch — Cursor · accessed 19 Aug 2026
“Agent uses grep automatically when you reference specific symbols.”

Practise 10 questions on this topic

Take Cursor AI Features — Timed Test 1 (10 questions) — scored instantly, explanation for every question, no login.

Start the test

Two developers chase the same bug. One asks the agent to look at whatever is wrong with the payment retry logic; the other writes that PaymentFailedError is thrown twice on retry inside PaymentService. The second run lands on the offending lines in a fraction of the steps. What accounts for the gap?

Related questions

Two developers chase the same bug. One asks the agent to look at whatever is wrong with the payment retry logic; the other writes that PaymentFailedError is thrown twice on retry inside PaymentService. The second run lands on the offending lines in a fraction of the steps. What accounts for the gap?

Related questions