How we cut GitLab Duo Code Review’s cost by 98%

The cost of the GitLab Duo Code Review AI agent was reduced by 98% while maintaining review quality. An initial 85% cost reduction was achieved by switching to a cheaper, open-weight model, but this was limited by large reviews exceeding the model's context window. These oversized reviews had to fall back to a more expensive model, capping the savings. The final 98% reduction was unlocked by implementing context compaction techniques, which made all reviews fit within the cheaper model's limits, demonstrating that both model selection and harness optimization were necessary.