The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains instructions that explicitly direct the agent to bypass human-in-the-loop safety protocols. Directives such as 'CONTINUE — do NOT stop to ask the user' and 'No human approval per iteration' remove the user's ability to review or intervene in the agent's autonomous actions.
[COMMAND_EXECUTION]: The autonomous loop relies on the 'Bash' tool to execute user-defined evaluation commands (e.g., 'uv run backtest.py') against code that the agent has just modified. This creates a feedback loop where the agent executes its own modified code without external verification, which is a high-risk dynamic execution pattern.

autoresearch