Search first, then read the source to cut agent tokens
For agent , top search snippets can be misleading if they are treated as proof. They are better seen as candidates that still need checking.
A stronger flow is to search first to narrow the options, then reopen the real source and read only the relevant part to confirm it. In one on a with about 2,000 files, a plain shell approach used 962 tokens on average and found 22 out of 24 targets.
A search-then-browse approach used 460 tokens on average and found 23 out of 24 targets. The result is limited because it used one agent and one corpus, but it suggests nearly half the token use with slightly better recall.
Key points
- Top search snippets should be treated as candidates, not final evidence.
- A two-step flow can reduce how much text an agent needs to read.
- In the small , search then browse used 460 tokens and hit 23/24 targets.
- The plain shell approach used 962 tokens and hit 22/24 targets.
- The result needs more testing because it used one agent and one corpus.