agentic-workflow

🔄 Agentic Workflow

Philosophy: Agents gather context, take action, verify, and iterate until the goal is achieved.

The Agentic Loop

Based on Anthropic's agent design patterns:

┌─────────────────────────────────────────────┐
│                                             │
│  1. GATHER CONTEXT                          │
│     └─ Read files, search, understand task  │
│                                             │
│  2. TAKE ACTION                             │
│     └─ Write code, run commands, edit files │
│                                             │
│  3. VERIFY RESULTS                          │
│     └─ Run tests, check output, validate    │
│                                             │
│  4. ITERATE OR COMPLETE                     │
│     └─ If not done, return to step 1        │
│                                             │
└─────────────────────────────────────────────┘

Core Principles

1. Start Simple

• Begin with a single-agent system.
• Add complexity only when needed.
• Maximize one agent's capabilities before adding more.

2. Clear Tool Contracts

Each tool should have:

• Clear input: What parameters it accepts
• Clear output: What it returns
• Error handling: What happens on failure

# Good tool definition
def search_codebase(query: str, file_pattern: str = "*") -> list[SearchResult]:
    """
    Search codebase for matching code.
    
    Args:
        query: Text or regex to search for
        file_pattern: Glob pattern to filter files (default: all)
    
    Returns:
        List of SearchResult with file, line, and content
    
    Raises:
        SearchError: If query is invalid regex
    """

3. Verification Before Completion

Never assume success. Always verify:

Action	Verification
File edit	Re-read file, check syntax
Code change	Run tests, check build
API call	Check response status
Command exec	Check exit code, read output