The abstract for Annotrace, a workbench for step-level annotation of agent traces, opens by noting a recurring problem: every study of why LLM agents fail begins by labelling traces by hand, and every such study builds its own throwaway tool. This suggests a lack of reusable infrastructure for this task.

The abstract also points out that observability platforms record agent traces but offer no structured labelling, while annotation tools label text—implying they are not designed for agent traces. Annotrace appears to be a response to this gap, though the abstract is truncated and does not describe the tool's features in detail.

Given the abstract's framing, Annotrace is positioned as a dedicated workbench for step-level annotation, potentially offering a structured alternative to ad-hoc tooling. However, the available text does not specify how it works or what it provides beyond the name and purpose.