# Replit Agent lets AI choose its own sub‑agents, cutting cost and boosting scores

> Replit’s new Agent lets a model decide when and how to delegate tasks, outperforming older setups on software‑engineering benchmarks.

Oossa · 2026-09-29 · https://oossa.com/en/replit-agent-lets-ai-choose-its-own-sub-agents-cutting-cost-and-boosting-scores

Replit released the Replit Agent, a system where the AI model itself picks sub‑agents, sets their size and effort, and decides when to reuse them. On the DeepSWE benchmark it scored 72 % for $2.11 per task, beating a side‑kick setup by 11 points. The same agent hit 49 % on Terminal‑Bench at $2.53, also ahead of the baseline. The change lets stronger models like GPT‑6 Astra delegate work to cheaper helpers, keeping expensive tokens for the hardest decisions.

## The facts

- DeepSWE score 72 % at $2.11 per task (Sept 2026)
- Terminal‑Bench score 49 % at $2.53 per task

## Why it matters

Letting the model control delegation improves performance while reducing compute costs for developers.

## Sources & references

1. [Free the models: Harness design at the frontier](https://replit.com/blog/free-the-models) – Replit, 2026-09-29

Last updated: 2026-09-29