The Agent Skills Directory

[DYNAMIC_EXECUTION]: Multiple scripts, such as scripts/generate_test_cases.py, scripts/compare_agents.py, and their JavaScript counterparts, use dynamic importing (importlib.import_module and import()) to load agent logic and graphs from local files. This is the primary mechanism for testing external code and is considered a standard functional requirement for this type of tool.
[EXTERNAL_DOWNLOADS]: The skill integrates with LangSmith (smith.langchain.com), a well-known service for LLM application evaluation. Scripts such as scripts/evaluate_with_langsmith.py fetch datasets and upload experiment results to LangSmith using the official SDK.
[INDIRECT_PROMPT_INJECTION]: As a tool designed to process agent trajectories and datasets, the skill has an inherent attack surface for indirect prompt injection. If an agent is tasked with evaluating a malicious dataset containing hidden instructions, those instructions could influence the agent's behavior. This is a low-severity risk typical of evaluation tools.
Ingestion points: scripts/run_trajectory_eval.py and scripts/evaluate_with_langsmith.py ingest data from local JSON files and remote LangSmith datasets.
Boundary markers: Not explicitly present; the scripts pass data directly to the agent or evaluator.
Capability inventory: scripts/compare_agents.py and scripts/run_trajectory_eval.py execute loaded agent functions.
Sanitization: Relies on standard SDK behavior and basic JSON parsing without specific sanitization of the input strings.
[COMMAND_EXECUTION]: The SKILL.md provides numerous examples of shell commands for the user to execute via uv run and node. These are intended for triggering the provided evaluation and testing scripts locally.

langgraph-testing-evaluation