Replit Agent beats sidekicks at lower cost

Replit Agent scored 72% on DeepSWE at $2.11 per task and 49% on Terminal-Bench at $2.53, beating its fixed-sidekick design by 11 and 16 points. In its production and benchmark report, Replit says GPT-6 Astra chose when to delegate and returned to existing subagents, while Astra alone reached higher scores only at more than twice Replit Agent’s cost. Builders should let capable models allocate subagents and effort dynamically, then measure cost per completed task—not copy a permanent-worker pattern—alongside Octobench’s harness-cost evidence.