general-quality

To evaluate agent traces for hallucination risk and response quality, you can score outputs against criteria like actionability, structure, and verified tool/API existence. This process checks runnable completeness and detects invented identifiers to ensure trustworthy responses.

Checking Claude Code outputs for actionability and structure involves verifying that responses contain complete, runnable code and concrete paths. It ensures clear organization with appropriate formatting like fenced code blocks and numbered steps for immediate user execution.

Detecting hallucinated APIs and deprecated Databricks features in agent outputs requires checking referenced tools against verified existence. This evaluation penalizes invented identifiers and confirms that all MCP tool references and API features are real and currently valid.

Error handling evaluation does check if execution failures are explained and recovered from in agent traces. It confirms that failures are addressed with recovery steps and penalizes unnecessary preamble or disclaimers that reduce response conciseness.

You can use response quality evaluation for any domain where an agent produces step-by-step outputs, code snippets, or tool/API references. It is especially effective when interacting with Databricks or MCP tools to verify runnable completeness and correct formatting.