Problem Statement
The workflow aims to generate a repeatable, rigorous evaluation framework for an LLM-based agent. It transforms inputs such as agent goals, tools, example conversations, success criteria, and failure modes into a structured rubric, representative test cases, scoring format, and analysis recommendations. The objective is to help builders benchmark agent performance both offline and during live monitoring without exposing sensitive data.
Steps
Ingest Project Inputs
INGESTCollect the agent goal, tool list, representative conversations or tasks, success criteria, known failure modes, and target model details. Parse these inputs into structured components used throughout the evaluation design.
Retrieve Domain Norms and Edge Cases
EXECUTEIdentify common evaluation patterns, adversarial behaviors, and domain-specific challenges. Pull in standard edge cases or known pitfalls in LLM agent evaluation to supplement user-provided details.
Generate Evaluation Rubric and Test Set
EXECUTEDraft a weighted evaluation rubric, construct 10–30 diverse test cases including edge and adversarial examples, and define expected outcomes. Ensure clarity on offline versus live monitoring components.
Validate Completeness and Rigor
EXECUTECheck that the rubric is balanced, test cases are representative, no sensitive data is included, and all success criteria are measurable. Confirm that scoring and analysis recommendations are feasible for builders.
Generate Scoring Sheet and Analysis Guidance
EXECUTEProduce a lightweight scoring format and instructions on interpreting results, identifying trends, and suggesting iteration paths for future improvements.
Prompt Configurations
Evaluation Framework Generation Prompt
You are an evaluation specialist tasked with designing a rigorous, builder-friendly framework to test an LLM agent. Use only the provided inputs: agent goal, tool list, example conversations or tasks, success criteria, known failure modes, and target models. Create: (1) a weighted evaluation rubric, (2) a representative test set with 10–30 cases that includes edge, adversarial, and standard cases, (3) a concise scoring sheet, and (4) analysis guidance separating offline evaluation from live monitoring. Do not invent tools or capabilities. Keep tone rigorous and experimental. Avoid any sensitive or private data.
Example Inputs/Outputs
Agent goal: answer user financial-planning questions using a predefined rule set. Tools: calculator API. Success criteria: accuracy, no hallucinated financial advice, rule compliance. Known failure modes: inventing tax rules, misinterpreting API errors. Target model: gpt-4o. Example conversation: user asks about retirement savings.
Rubric with dimensions such as rule compliance (30%), accuracy (40%), tool usage correctness (20%), and clarity (10%). 12 test cases including: user asks for restricted tax advice (expect refusal), calculator API returns error (expect graceful fallback), adversarial user pushing for forbidden recommendations. Scoring sheet with 1–5 per dimension. Analysis guidance notes whether failures cluster around tool handling or rule interpretation, and recommends tuning refusal behavior and error-handling prompts.