The ecosystem surrounding Large Language Model (LLM) agents is expanding rapidly. Today, developers are no longer just writing monolithic codebases; instead, they are building agentic workflows using modular, reusable components known as "skills."

These skills are mixed-modality packages containing natural language instructions, code fragments, API tools, metadata, and specific operational workflows. As marketplaces for these agent skills emerge, a new security and intellectual property challenge has surfaced: how do we audit and track the reuse of these multi-modal assets?

Traditional code clone detection methods are ill-equipped to solve this problem. Enter SkillTrace, a multi-trace provenance auditing framework designed specifically for LLM-agent skill reuse.

The Challenge: Why Traditional Audits Fail

Historically, software auditing relied on detecting code similarity. If Developer B copied a block of code from Developer A, standard AST (Abstract Syntax Tree) matching or token-based scanners could flag the duplication.

However, LLM-agent skills are fundamentally different. A skill is not just code; it is a blend of:

• Expression: Natural language prompts and system instructions.
• Implementation: Code snippets, tool definitions, and API integrations.
• Operation: The logical flow of execution, decision-making logic, and external resource activation.

If someone copies a skill but rewrites the python implementation while keeping the exact prompt structure and tool execution flow, traditional detectors will miss it. Similarly, if they rewrite the prompts but keep the exact underlying operational graph, it still constitutes skill reuse.

Introducing SkillTrace

To address this gap, researchers Jialuo Chen, Minghe Wang, and their co-authors introduced SkillTrace in their paper, "SkillTrace: Multi-Trace Provenance Auditing for LLM-Agent Skill Reuse" (arXiv:2608.05204).

SkillTrace breaks down the auditing process by extracting and analyzing three distinct "provenance traces":

1. The Expression Trace

This trace captures the linguistic footprint of the skill. It analyzes natural-language prompts, developer-written metadata, and system instructions. Even if the underlying code changes, similarity in the expressive logic can be surfaced here.

2. The Implementation Trace

This trace focuses on the software engineering components of the skill. It audits code fragments, API hooks, and tool references, identifying instances where structural implementation logic has been copied or slightly modified.

3. The Operational Trace (Skill Operational Graph)

Perhaps the most innovative aspect of SkillTrace is the Skill Operational Graph (SOG). The SOG maps the dynamic structure of the skill, outlining activation paths, procedures, and the flow of resources. Because extracting this structure from unstructured files is complex, SkillTrace utilizes an LLM to assist in generating the SOG once, at the time of ingestion.

Designing for Speed and Scale

Using LLMs to evaluate similarity at scale can be prohibitively expensive and slow. To solve this, SkillTrace adopts a hybrid architecture:

• LLM-Assisted Ingestion: The LLM is only utilized during the initial ingestion phase to help build the Skill Operational Graph (SOG).
• Deterministic Auditing: At audit time, SkillTrace compares cached traces deterministically. This removes the latency and cost of calling an LLM API during the comparison phase.
• Calibration: The system calibrates trace scores against "same-function strict negatives" (skills that perform the same task but were written completely independently) to minimize false positives.

Proven Performance

To evaluate the system, the researchers built SKILLTRACE-BENCH, a benchmark consisting of 820 transformed reuse positives across 100 marketplace anchors, alongside 751 negative controls.

In testing, SkillTrace achieved remarkable results:

• AUROC: 0.938
• F1-Score: 0.898

Furthermore, when deployed in a "wild audit" of 36,446 real-world skills, SkillTrace successfully surfaced actionable review queues, proving that it can scale to meet the demands of commercial agent registries and open-source repositories.

Why This Matters for the Future of AI

As autonomous agents take over workflows from customer support to software engineering, modular skill marketplaces will become high-value targets for intellectual property theft and unauthorized reuse. Tools like SkillTrace ensure that developers can protect their creative prompts and complex operational logic, paving the way for a more secure, transparent, and fair developer ecosystem in the era of agentic AI.