# Replit — Free the models: Harness design at the frontier

- Company: Replit (replit.com)
- Announced: 2026-09-29T16:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://replit.com/blog/free-the-models
- Record: https://forck.live/items/17881-free-the-models-harness-design-at-the-frontier
- Subject: Replit Agent
- Models affected: GPT-6 Astra

Replit describes how its Replit Agent lets the main model choose its subagents' tier and effort instead of relying on a router, arguing that a router is always less capable than the model it selects for. The company reports that on DeepSWE and Terminal-Bench, Replit Agent is Pareto-efficient against Astra on its own and beats a sidekick architecture by 11 and 16 points.

## Evidence

Verbatim from https://replit.com/blog/free-the-models:

> Given that freedom, GPT-6 Astra hands routine implementation to less costly subagents and decides for itself where its tokens are worth spending.

---

Record: https://forck.live/items/17881-free-the-models-harness-design-at-the-frontier
Catalogue: https://forck.live/llms.txt
Current issue: https://forck.live/feed.md