Agent Infrastructure, Harness Design, and Multi-Model Economics

Recent analysis emphasizes that cheaper models do not necessarily reduce agent bills: task success, verification, workflow design, routing, caching, and tool-call control can produce large cost differences, reportedly up to 30x between runs. Typed workflow DSLs, persisted checkpoints, tool validation, observability, resumable execution, and agent-driven performance tooling remain important complements to larger models and context windows.