Free the models: Harness design at the frontier

Replit describes an agent architecture that lets a core language model dynamically decide its own reasoning effort, whether to delegate work, and which specialized subagents to use, rather than relying on a fixed external router or a single persistent sidekick. The system provides domain-specific workers, multiple cost and capability tiers, reusable warm subagents, and mid-turn effort adjustment where model caching permits, aiming to match resources to task complexity and exploit different models’ strengths in coding, design, testing, and writing. Production observations suggest that newer frontier models, particularly GPT-6 Astra, increasingly choose to delegate routine work and revisit previously briefed workers while retaining difficult decisions themselves. In reported evaluations on DeepSWE and Terminal-Bench, Replit Agent using Astra achieved a more favorable score-cost tradeoff than published Astra-only baselines and substantially outperformed a single-sidekick configuration, although Astra at higher spending achieved higher absolute scores. The authors frame the approach through Sutton’s “bitter lesson,” arguing that increasingly capable models benefit from flexible, composable harnesses that leave execution choices to the model rather than embedding rigid human-designed workflows.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.