The Challenge of Skill Provenance in Agentic Workflows

As LLM agents become more sophisticated, they increasingly rely on 'skill reuse'—the ability to leverage previously learned or stored procedures to solve new tasks. However, this capability introduces a significant 'black box' problem: when an agent executes a complex task using a library of skills, it becomes difficult to trace which specific prior experiences or training traces informed the current decision-making process. SkillTrace addresses this by implementing a multi-trace provenance auditing system, allowing developers and researchers to map the lineage of agentic actions back to their source traces.

Multi-Trace Auditing Architecture

SkillTrace functions by capturing and indexing execution traces during the agent's learning phase. When an agent attempts to reuse a skill, the system performs a provenance check that correlates the current context with the historical data that originally generated the skill. This allows for:

• Attribution: Identifying which specific past interaction or training data point contributed to a particular skill's current implementation.
• Verification: Ensuring that the reused skill is being applied in a context consistent with its original intent, reducing the risk of 'hallucinated' or misapplied capabilities.
• Debugging: Providing a clear audit trail when an agent fails, enabling developers to see if the error originated from the base skill or the current execution context.

By moving beyond simple skill storage to a provenance-aware architecture, SkillTrace provides a necessary layer of accountability for autonomous systems operating in complex, multi-step environments.