The Agent Skills Directory

Command Execution via Harbor Framework: The skill utilizes the Harbor CLI (harbor task init, harbor run) to manage and execute agent evaluation tasks. It specifically recommends using Docker containers for environment isolation, which is a standard security practice for running agent code in a controlled and restricted manner.
External Data Sourcing from LangSmith: The skill provides instructions for using the LangSmith CLI to fetch agent traces. This integration allows developers to use real-world interaction data to design more effective evaluations. LangSmith is a well-known observability platform provided by the skill's authoring organization.
Credential Management Best Practices: The documentation explicitly instructs users to avoid hardcoding secrets and instead pass necessary credentials via environment variable references at runtime. It also warns against including sensitive information in fixtures or logs, aligning with industry security standards.
Mitigation of Indirect Prompt Injection: The skill includes design patterns for verifiers that help protect against indirect prompt injection. It instructs the use of delimiters and specific prompts that tell LLM judges to ignore potential instructions embedded within the target agent's output, thereby reducing the risk of the agent manipulating its own evaluation results.

eval-engineering