agentsop-output-format-by-model

Choosing the right structured output format involves matching the wire format to your content and consumer. Select between plain-text with diff, markdown, typed JSON fields, or tool_use envelopes based on whether a parser, human, or tool dispatcher consumes the output next.

Use plain-text with diffs for code edits to avoid nesting code inside JSON, which degrades quality. Apply JSON typed fields or tool_use envelopes with scalar arguments when a downstream tool dispatcher requires structured extraction rather than raw code generation.

Decide the output format by evaluating the generated content and its downstream consumer. Follow an empirical workflow to choose between markdown, JSON tool_use envelopes, or grammar-constrained output, ensuring the wire format matches the target model's strengths.

Forcing structured JSON outputs can limit reasoning ability due to format-restricted constraints. Grammar-constrained output or markdown often yields better results for complex generation tasks, while JSON should be reserved for typed field extraction and tool-call payloads.

Grammar-constrained output is applicable for RAG and extraction pipelines to enforce schema compliance. It provides an alternative to JSON typed fields when you need strict format adherence without imposing the wire-format tax that degrades model performance.