The Decomposition Is the Fingerprint: Per-Component Identity for Agent Skills

As AI agents increasingly rely on reusable skills—bundles of prompts, code, and tool declarations—governance teams face a critical identity problem: how to deduplicate, version, and audit these skills when they are frequently modified. Traditional cryptographic hashes fail because they are too sensitive to minor edits, while raw embedding comparisons are computationally expensive and lack interpretability. The authors propose a per-component, locality-sensitive fingerprint that embeds each skill component (prompt, code, tools) and projects it into a 320-bit multi-bank SimHash. This results in a fixed 120-byte signature that allows for constant-time comparison using Hamming distance.

• High Fidelity: The 120-byte signature reproduces the ranking of a 24,576-bit embedding to within 1% AUC, achieving an overall AUC of 0.974 across 4,950 pairwise comparisons.
• Component-Level Insight: By maintaining a triple of component scores rather than a single scalar, the system can distinguish between different types of reuse, such as repurposing (same code, different prompt) versus cloning (both components match).
• Adversarial Robustness: The fingerprint survives aggressive rewrites, including cross-language translation and full paraphrasing, where lexical baselines (like TF-IDF) collapse. It correctly identifies tampered skills as near-clones of a known base and localizes the change to the specific component.
• Registry Utility: The fingerprint enables efficient registry operations, including family clustering, novelty detection, and the generation of a portable "SkillBOM" (Bill of Materials), all while maintaining O(1) comparison costs.

This work provides a foundational primitive for governing agent ecosystems. By separating structural identity from behavioral integrity, it allows platforms to build registries that can track the lineage of agent skills without sacrificing performance. It demonstrates that identity is not a single number, but a structured relationship that can be managed through compact, bit-level signatures.

AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.