The Agent Skills Directory

[PERSISTENCE]: The skill utilizes CronCreate to schedule recurring autonomous tasks that execute instructions without immediate user oversight. These scheduled jobs persist for up to 3 days.
[COMMAND_EXECUTION]: The autonomous loop involves the execution of system commands, including git checkout, git add, git commit, and python script invocation within the agent's context.
[DYNAMIC_EXECUTION]: The skill creates a self-modifying execution cycle. The agent is instructed to autonomously edit a target file and then immediately execute a Python-based evaluation script (run_experiment.py) to test the changes.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a significant attack surface as its recurring behavior is dictated by the contents of files within the workspace.
Ingestion points: The agent reads strategy, constraints, and evaluation commands from .autoresearch/{domain}/{name}/config.cfg and program.md.
Boundary markers: The instructions do not define clear boundaries or validation steps for the content ingested from these files.
Capability inventory: The loop has the capability to write files, commit changes to the repository, and execute shell commands.
Sanitization: There is no evidence of sanitization or integrity checking for the configuration files before they are used to influence the agent's autonomous actions.

loop