PHANTOMTRACE: Finding the Ghosts Inside AI Agent Execution Traces 👻

Описание и применение

PHANTOMTRACE позволяет разработчикам быстро выявлять и понимать проблемные шаблоны в трассах выполнения AI‑агентов, улучшая прозрачность отладки.

Задача

AI agents often repeat actions, encounter errors, or retry failed operations, making raw execution logs hard to interpret.

Результат

PHANTOMTRACE enables developers to quickly identify and understand problematic patterns in AI agent execution traces, improving debugging transparency.

Рабочий процесс

• 1. Collect AI agent execution logs in JSONL format.
• 2. Run PHANTOMTRACE to parse and analyze the logs.
• 3. Detect patterns such as repeated tool calls, multiple failures, and immediate retries using Python rules.
• 4. Visualize the results in a Streamlit dashboard with Plotly charts.
• 5. (Optional) Use a local Ollama/Gemma 3 1B model to generate explanations for detected patterns.

Доказательства

PHANTOMTRACE helps developers investigate these execution traces and identify suspicious patterns.

Источник

It identifies patterns such as: Repeated tool calls. Multiple failed operations. Immediate retries after failures.

Источник