Tuning Deep Agents to Work Well with Different Models

Deep Agents has introduced model-specific profiles to optimize prompts, tools, and middleware for different model families, such as OpenAI, Anthropic, and Google models, enhancing performance by 10-20 points on the tau2-bench. Previously, Deep Agents operated with a generic set of parameters that worked across all Large Language Models, but the new harness profiles allow for tailored adjustments, improving efficiency and adherence to model-specific prompting guides, such as OpenAI's Codex and Anthropic's Claude. This customization is vital because different models respond differently to specific prompting techniques, as demonstrated by improved results on Terminal-Bench 2.0, where adjustments elevated gpt-5.2-codex from 52.8% to 66.5%. The profiles are declared as override layers for system prompts, tool inclusion, and naming, among other factors, and can be registered or overridden by users to suit specific needs, while maintaining consistent call sites. The introduction of these profiles empowers builders to manage and test agent configurations effectively, with the goal of creating optimal harnesses for various tasks.