Traceburn, a local profiler that found 69% avoidable agent spend (github.com)

Traceburn has launched as a local profiler designed to enhance cost-efficiency in AI agent operations, claiming to identify a substantial 69% of avoidable spending in the execution of static policy prompts. In a demonstration, a support ticket triage agent executed ten calls with a static prompt of 4,700 tokens, costing $0.0539 uncached. By implementing Traceburn's suggested cache control block, the cost was reduced to $0.0167. Traceburn's efficient methodology features a waste report that quantifies unnecessary spend, providing developers an actionable way to optimize costs without requiring extensive reconfiguration. Significantly, Traceburn operates entirely locally with no external telemetry or user accounts needed—keeping all data private and secure. It leverages Python SDKs from OpenAI and Anthropic to create a detailed cost and latency flamegraph, alongside deterministic replay capabilities for recorded calls. Additionally, Traceburn introduces a unique waste detection mechanism that checks for inefficiencies like duplicate calls and excessive prompts. This localized approach not only aids developers in identifying and minimizing expenditures but also fills a gap in the existing landscape of AI profiling tools by prioritizing ease of use, precision in assessing waste, and maintaining privacy.