Evaluating agents and their limits
WebAI evaluation means testing systems against defined tasks and evidence rather than equating fluency with reliability. Models, retrievers, agents and interfaces need different metrics and appropriate human review.
Define the evaluation target
An application can fail during document retrieval, answer generation, tool selection, approval or final display. Identify components and objectives before choosing metrics. A global success rate does not reveal which step actually failed.
Test sets and reproducibility
Prepare representative, negative, ambiguous and contradictory cases with reference answers or evidence. Keep model version, date, parameters, data and annotation methods. A small demo set does not establish robust general performance.
Evidence faithfulness
In RAG, measure document retrieval, passage relevance, answer faithfulness and citation validity separately. One supported statement may sit beside unsupported statements; evaluate claims individually.
Evaluate acting agents
An agent may write correct prose but perform the wrong operation. Review permissions, execution traces, tool calls, approvals and actual final state. Test environments should isolate sensitive actions and permit cancellation.
Metrics and human judgment
Relevance, accessibility, bias, privacy, security and educational quality cannot be captured by a single automated score. Combine technical checks, human review, independent assessment and error analysis. Record disagreement rather than hiding it.
Exercise: audit an assistant
Build twenty questions with verified sources, five unanswerable cases and safe adversarial inputs. Compare two setups, classify failures and explain sample limitations. Do not publish only a success rate; show difficult examples too.
Core steps
- 01Decompose the system and its risks
- 02Build a documented test set
- 03Measure components separately
- 04Test negative and adversarial cases
- 05Request independent human review
- 06Publish failures, limitations and versions
Three questions to ask
Does fluent writing establish accuracy?
Was the final action state checked?
Are difficult failures reported?
Is the evaluation corpus representative?
Read the source documents
Versions, reuse terms and document status can change. Check them on the original sites.