The Cursor team released the "Agent Trace" specification (version 0.1.0, status as RFC) in agent-trace.dev, positioning it as an open specification for documenting "which changes come from AI and which are from humans" in a versioned codebase, and correlating model information with relevant conversations to specific code contributions.
The specification emphasizes vendor neutrality and interoperability, supports documenting attribution data at file and row granularity, and allows for expanded fields to be compatible with different tools. The page also clarifies that the specification is not used to determine legal code ownership or copyright ownership, nor is it responsible for evaluating the quality of AI contributions. At its core, it provides a readable, exchangeable record format for different development tools to write and read.
FAQs
Q: Which team proposed the specification for Agent Trace?
A: Agent Trace was proposed by the Cursor team and released on agent-trace.dev.
Q: What problems does Agent Trace mainly solve?
A: It is used to track the origin and attribution of AI-generated code in the codebase, distinguishing between human and AI contributions.
Q: Is Agent Trace a product or standard?
A: It is a data specification rather than a specific product that defines how attribution data is recorded.
Q: Does Agent Trace handle copyright attribution or ownership?
A: The page states that the specification is not used to determine ownership or copyright issues at the legal level.
Q: How granular attribution granularity does Agent Trace support?
A: Objectives support file-level and row-level attribution records and extend metadata fields.
