Deep Agent Tuning for Different Models

Deep Agent previously provided a single harness (prompts, tools, middleware) that worked across all LLMs. However, prompting guides differ by model, and even the same model's performance could vary significantly depending on the harness configuration.

To address this, LangChain has released the harness profiles feature. It allows customizing system prompts, tool implementations, and middleware for each model, and provides default profiles for OpenAI, Anthropic, and Google models.

Performance improvement results (based on tau2-bench subset):

• GPT 5.3 Codex: 33% → 53% (20 point improvement)
• Claude Opus 4.7: 43% → 53% (10 point improvement)

Key changes:

Codex profile:

• Uses the recommended tool apply_patch instead of the default edit_file
• Aliases the execute tool to shell_command
• Added prompting related to parallel tool call batch processing

Claude Opus profile:

• Added reflection guidance on tool results (<tool_result_reflection>)
• Encourages checking state directly via tools instead of memory (<tool_usage>)

Profiles can be defined in Python code or YAML, and are automatically applied when calling create_deep_agent(model="..."). Custom profiles can also be registered or distributed as plugins.