managing-agentic-loops

The Agentic Loops skill provides the procedural 'how-to' for implementing autonomous reasoning cycles. It is designed for use by Operational Intelligence agents to navigate complex, non-linear tactical objectives with high reliability.

When to Use

This skill should be used when completing tasks related to agentic loops.

Process

Follow these procedures to implement the capability:

1. The ReAct Core (Reason → Act → Observe)

The foundational pattern for any autonomous agent.

• Reasoning Phase: The agent must state its current understanding, its intent, and the reasoning behind its tool choice.
• Action Phase: Precise execution of a single tool call.
• Observation Phase: Objective recording of the tool's output without immediate interpretation.
• Loop Termination: The cycle ends only when the goal is met or a max_iterations fail-safe is triggered.

2. Multi-Dimensional Thinking (Sequential Thinking)

For complex problem solving, integrate the sequential-thinking MCP tool directly into the loop.

# Sequential Thinking Integration Pattern
async def complex_problem_solver(objective: str):
    thinking_process = await call_tool("sequential-thinking", {
        "thought": f"Decomposing the objective: {objective}",
        "thoughtNumber": 1,
        "totalThoughts": 5,
        "nextThoughtNeeded": True
    })
    # Use output to drive the next ReAct iteration

3. Reflective Quality Gates

Implement a secondary reflection pass to evaluate and correct the agent's own output.

• Reflection Prompting: "Evaluate your last 3 actions against the Root Goal. Did you advance the strategy or are you stalling? If stalling, propose a radical change in approach."
• Self-Correction Logic: Inhibit tool execution if the reflection score is below a predefined threshold (e.g., 0.8 consistency).

Process

| Objective | Preferred Sequence | | : | : | | Code Refactoring | view_file → codebase_search → sequential-thinking → replace_file_content → pytest | | Structural Audit | list_dir → find_by_name → grep_search → link_checker | | New Feature Dev | read_url_content (Specs) → implementation_plan (Doc) → task_boundary → write_to_file |

Best Practices

1. Strict State Management: Always update task.md or a state registry between major loop transitions to prevent context fragmentation.
2. Explicit Tool Intent: The agent must declare its intent before calling a tool (e.g., "I will now list the directory to verify the path exists").
3. Fail-State Awareness: Explicitly handle common errors (e.g., FileNotFoundError, API Timeout) with retry-backoff or human-in-the-loop (HITL) triggers.

Anti-Patterns

• Hallucinated Tool Args: Calling tools with parameters that don't exist in the definition.
• Reaching 'Max Iterations' without Observation: Continuously reasoning without taking actions or observing results.
• Ambiguous Reasoning: Providing generic "I will do X" without explaining why X is the correct next step.

Related Entities

• Specialist: CognitiveCycleEngineer (Primary User)
• Knowledge: agentic-loop-patterns.json
• Workflow: Standard Feature Delivery Cycle (SFDC)

Prerequisites

• Access to relevant project documentation
• Environmental awareness of the target stack

Best Practices

• Follow the system axioms (A1-A5)
• Ensure all changes are verifiable
• Document complex logic for future maintenance