shell-agent-delegation

Shell Agent Delegation for Resilient Workflow Execution

When to Use This Skill

Apply this pattern when:

• Direct tool execution (execute_code_sandbox, read_webpage, search_web) fails with 'unknown error'
• Multiple tool attempts have failed in sequence
• The task requires complex document generation or data processing
• You need a tool that can autonomously select libraries and handle multi-step workflows

Why This Works

The shell_agent tool differs from direct execution tools in key ways:

• Autonomous tool selection: Decides whether to use Python or Bash based on the task
• Built-in error recovery: Automatically retries and fixes errors (up to several rounds)
• Iterative execution: Writes code, executes, inspects output, and adapts
• Full workflow ownership: Handles the entire task end-to-end without manual intervention

Step-by-Step Instructions

Step 1: Recognize the Failure Pattern

Identify when to pivot to shell_agent:

- execute_code_sandbox returned 'unknown error'
- read_webpage/search_web failed multiple times
- Direct approaches are struggling with the task complexity

Step 2: Formulate the Delegation Task

Create a clear, self-contained task description for shell_agent:

Good task description:

Create a 1-page SBAR Template PDF document. Include sections for:
- Situation: Brief description of the current situation
- Background: Relevant context and history
- Assessment: Current assessment and analysis
- Recommendation: Proposed actions and next steps
Use a professional layout with clear headings and adequate whitespace.

Key elements to include:

• The end goal (what should be produced)
• Required sections/components
• Format requirements (PDF, DOCX, etc.)
• Any style or layout preferences

Step 3: Execute the Delegation

Call shell_agent with your task description:

# Conceptual example
shell_agent(task="Create a professional SBAR Template PDF with Situation, Background, Assessment, and Recommendation sections. Include clear headings and professional formatting.")