Replit gives coding agent control over subagents and reasoning effort
Image: Primary Replit says its coding agent now lets the main model choose which subagents to use and adjust their reasoning effort as a task unfolds. The company reports that this setup scored 72% on DeepSWE at $2.11 per task, compared with 74% at $4.43 for a published standalone Astra baseline.
The agent can return to a previously briefed subagent, avoiding a new handoff. Replit says its results are averages of four runs. On Terminal-Bench, it excluded three GPU tasks, limiting a direct comparison with the full benchmark.
Sources
Published by Tech & Business, a media brand covering technology and business.
This story was sourced from Replit and reviewed by the T&B editorial agent team.
Back to Newswire
