⬡ AgentLens

Failure Cluster Analysis

🔬 How clustering works: Failed agent outputs are embedded using sentence-transformers and grouped with HDBSCAN — then each cluster is auto-labeled by an LLM. The clusters below were computed locally and pre-loaded into the dashboard. To refresh clusters, run python3 evaluator/clustering.py locally and redeploy.
Clusters Found
5
Runs Clustered
15
Avg Cluster Score
0.813
2
Fails by outputting fragments or keywords instead of complete sentences.
all 4 runs avg score: 0.734
Sample Outputs
"le texte en français"
"text, phrase, substantive information, narrative, data"
"Paris"
0
Fails by outputting only a single number or word.
all 3 runs avg score: 0.812
Sample Outputs
"1200"
"1200"
"Zero"
3
Fails when empty or invalid input triggers JSON parse errors.
all 3 runs avg score: 0.833
Sample Outputs
"Agent could not complete the task within the step limit."
"The provided text is empty, so there is no content to summarize. Please provide the text you would like me to condense i..."
"Pipeline failed. Errors: Tool 'parse_json' failed: JSON parse error: unexpected token at position 0 in: ' AgentLens is a..."
4
Confuses general LLM concepts with specific tool capabilities.
all 3 runs avg score: 0.796
Sample Outputs
"LLM observability, agent pipelines, execution traces, regression detection, FastAPI"
"The text outlines key components of an LLM operations (LLMOps) ecosystem, focusing on agent pipelines and observability ..."
"Based on the provided context, AgentLens currently supports **Groq-hosted models**, including **qwen/qwen3.8-27b**. Addi..."
1
Refuses to answer due to insufficient document information.
all 2 runs avg score: 0.889
Sample Outputs
"I don't have enough information in the provided documents."
"I don't have enough information in the provided documents."