The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill implements a meta-optimization loop in meta_harness.py that relies on a proposer agent to generate new harness configurations and code modifications. These LLM-generated artifacts are then parsed and executed during the evaluation phase to determine their efficacy.
[INDIRECT_PROMPT_INJECTION]: The skill architecture is designed to ingest and process external datasets for evaluation, which presents a vulnerability to instructions embedded within that data.
Ingestion points: Data enters the system through the train_dataset and val_dataset arguments in the meta_harness_loop function within meta_harness.py.
Boundary markers: The provided prompt construction in build_proposer_prompt (claude_wrapper.py) lacks explicit markers to differentiate between system instructions and potentially malicious content within the history or datasets.
Capability inventory: The framework utilizes the Anthropic SDK for network communication, includes capability for system interaction via subprocess, and executes evolved code logic during evaluation steps.
Sanitization: There is no evidence of input validation, sanitization, or filtering of the dataset content before it is processed by the harness.

meta-harness-optimization