What flat scores hide
A stable evaluation score can conceal a regression. Inspect the cases behind the number before deciding a change is ready.
Read articleThe Agentagon blog
Better agents.
Measured, not assumed.
Notes on understanding agent behavior, finding regressions, and proving that a change helps.
Observability vs evaluation for AI agents
Traces record what happened. Evaluations test expectations. Improving an agent needs a deliberate connection between the two.
Read articleAI agent regression testing beyond pass or fail
Turn a prompt, model, or tool change into a repeatable comparison with explicit expectations and inspectable evidence.
Read article