The Agent Skills Directory

[PERSISTENCE]: The skill implements a persistence mechanism by using the CronCreate tool to schedule recurring autonomous tasks at intervals ranging from 10 minutes to one month. This ensures the agent continues to perform actions without direct user initiation.
[COMMAND_EXECUTION]: The skill executes multiple shell commands as part of its automated loop, including git checkout, git add, git commit, and python {skill_path}/scripts/run_experiment.py. These commands run automatically at the scheduled intervals determined by the cron configuration.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because the autonomous loop's behavior is guided by data ingested from local files.
Ingestion points: The agent reads strategy, constraints, and metrics from .autoresearch/{domain}/{name}/config.cfg, program.md, and results.tsv.
Boundary markers: The prompt template lacks boundary markers or instructions to treat data from these files as untrusted, potentially allowing instructions embedded in those files to influence the agent's logic.
Capability inventory: The agent has the capability to write to the filesystem, perform git operations, execute python scripts, and manage its own persistence (creating and deleting cron jobs).
Sanitization: There is no evidence of sanitization or validation of the input read from the experiment files before it is processed by the agent.
[DYNAMIC_EXECUTION]: The skill facilitates dynamic code modification and execution. It specifically instructs the agent to edit a target file with new code changes and then execute an evaluation script (run_experiment.py) to test the resulting code.

loop