Víctor Sánchez BelmarResearch, field notes, experiments and explanations.https://victorsb.com/enThe $2.82 plan that made a local 27B model beat every frontier agent I testedhttps://victorsb.com/blog/local-27b-model-planning/https://victorsb.com/blog/local-27b-model-planning/Same Qwen model, harness and machine. An 11,044-token execution plan moved the score from 68.57 to 90.75, and all three blind reviewers picked it first.Tue, 18 Aug 2026 00:00:00 GMTlocal AIcoding agentsbenchmarks