The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to dynamically modify a training script (train.py) and execute it using uv run or bash. This pattern of runtime code modification and execution is a core feature of the automated experimentation loop.
[EXTERNAL_DOWNLOADS]: The skill uses uv run to execute scripts, which may involve downloading official packages such as numpy from well-known registries.
[PROMPT_INJECTION]: The skill contains an attack surface for indirect prompt injection via the files it processes to determine its actions.
Ingestion points: The agent reads EXPERIMENT.md and results.tsv to decide on the next experiment iteration.
Boundary markers: No specific delimiters or safety instructions are provided to help the agent distinguish between data and potentially malicious instructions within these tracking files.
Capability inventory: The agent is equipped with Edit and Bash tools, allowing it to modify and run the training code.
Sanitization: There is no evidence of validation or sanitization for the inputs retrieved from the experiment files before they influence the agent's code generation.

autoexperiment