The open-source tool Wattage analyzes AI agent traces to identify token waste, calculate associated costs, and provide actionable fixes for optimizing LLM performance and reducing operational expenses.

Key Points

• Wattage processes OpenTelemetry (OTel) JSON traces to detect inefficiencies like redundant tool calls, prompt prefix churn, and non-convergent loops.
• The tool provides a 0–100 token efficiency score and generates visual reports, including HTML flame graphs and status badges for project documentation.
• A built-in CI integration allows developers to set cost-regression gates that fail builds if agent changes exceed defined price or efficiency thresholds.
• The software operates fully offline without requiring API keys, ensuring data privacy while analyzing agent behavior and token consumption patterns.
• Its convergence engine identifies complex thrashing patterns that standard exact-match detectors often miss, achieving high precision in identifying stalled agent loops.

Why it Matters

As AI agents become more complex, hidden inefficiencies in token usage can lead to significant, unexpected cloud infrastructure costs. Wattage provides a standardized, automated way for engineering teams to monitor and gate these expenses, ensuring that performance optimizations are maintained throughout the development lifecycle.