Can you recommend a prompt testing tool for comparing agent behavior across structured output workflows?

Yes — if you’re comparing agent behavior across structured output workflows, I’d recommend looking at LangSmith first.

Why LangSmith

• Great for prompt/version comparison: you can run the same workflow across prompt variants and inspect outputs side by side.
• Structured output friendly: works well when your agents return JSON, Pydantic models, tool-call traces, or other schema-constrained outputs.
• Tracing for agents: lets you see intermediate steps, tool calls, retries, and failures — useful when behavior changes aren’t just in the final output.
• Evaluation support: you can define custom evaluators for schema validity, field accuracy, consistency, etc.
• Dataset-based testing: good for running the same structured cases across models/prompts.

Other strong options

• OpenAI Evals
  Best if you want a more evaluation-centric workflow and are already heavily in the OpenAI ecosystem. Less focused on interactive agent tracing than LangSmith.

• Promptfoo
  Very good for prompt regression testing and comparing outputs across models. Especially useful for structured outputs because you can assert on JSON fields, schema compliance, and diffs. Lighter-weight than LangSmith.

• Humanloop
  Good for prompt experimentation and evaluation with collaboration features. Useful if you want non-engineers involved in review.

• Weights & Biases Weave
  Solid if you already use W&B and want experiment tracking plus traces for LLM applications.

My practical recommendation

• If you want deep agent tracing + structured workflow debugging: LangSmith
• If you want simple regression tests and comparison across prompts/models: Promptfoo
• If you want formal evaluation harnesses: OpenAI Evals

If you want, I can also give you a side-by-side comparison table of LangSmith vs Promptfoo vs OpenAI Evals specifically for structured JSON outputs and agent workflows.