Sweet-search combines local retrieval with agent instructions

Account for the whole intervention

Setup changes agent instructions as well as search tools. The published task benchmark uses version 2.9.0 on 200 tasks across five model-and-harness configurations. Its cost figures are token-based list-price equivalents from subscription runs, rather than invoices.

Sweet-search combines local retrieval with agent instructions ↗

Keep accuracy beside cost

The author reports the same pooled solve count across the two arms, with differences by configuration. A separate retrieval evaluation reports a model-specific accuracy regression linked to following the search rules too literally. These findings do not establish equivalent behavior on every repository.

Sweet-search combines local retrieval with agent instructions ↗

What you can do

1. Inspect the per-configuration results and setup changes before a controlled trial. Open reference ↗

Where the evidence stops

Repository documentation read November 1; no code or benchmark executed. Independent replication and an original video demonstration were not found.

Sources read. Primary sources read November 1, 2026 UTC. No third-party installation, inference or complete workflow executed.

Sources & further reading