September 30, 2026
Replit Agent chooses a helper for each task: 72% on DeepSWE
On Sep 29, Replit Agent scored 72% on DeepSWE v1.1 in Max mode with GPT-6 Astra, at $2.11 per task. The primary agent adjusts its own effort level and selects one of three levels for a subagent, each with a separate Effort setting. Replit tested the product without prompt or wrapper changes.
Replit ⠕
@replit
Models are getting smarter. But the wrapper is not going anywhere. Now it does not make decisions for the model; it gives it options. In Replit Agent, the primary agent decides for itself when to think longer, when to delegate work to a subagent, and how much effort to allocate to it.
· 1.7K views
Previously, the wrapper made decisions for the model. Now Replit Agent decides for itself when to think longer and when to delegate part of the work to a subagent.
Tasks are split by weight. Replit sends mechanical name replacements to small/low. A complex bug that requires hypotheses gets large/high.
A Core or Pro user can select the primary model in Agent Mode for the next message by enabling Power or Max.
In Free Mode, Replit always selects the model itself, while manual selection is available to Core and Pro users in Power or Max.
Source
