Agentic Prompting (Planning, Reflection, Tool Use)

Simple Definition:

Agentic prompting is a set of techniques that instruct an LLM to behave like an autonomous agent - breaking down tasks into steps, reflecting on its own reasoning, using tools when needed, and iterating until the task is complete. Instead of one-shot answers, the LLM operates in a loop.

Think of the difference between asking someone "What is 2+2?" (simple prompting) versus saying "Here is a complex math problem. Break it down, solve each part, check your work, and correct any mistakes" (agentic prompting). Same person, but the instructions unlock a completely different mode of thinking.

Real-World Analogy - IAS Officer vs Clerk:

A government clerk follows exact orders - "stamp this file, forward it." An IAS officer gets a goal - "solve the water crisis in this district" - and then plans, delegates, reviews progress, course-corrects, and sees it through. Agentic prompting turns the LLM from a clerk into an IAS officer.

The Three Pillars of Agentic Prompting:

• Planning: Breaking a complex goal into actionable sub-tasks with dependencies
• Reflection: Self-evaluating output quality and reasoning correctness
• Tool Use: Deciding when internal knowledge is insufficient and external tools are needed

Agentic vs Traditional Prompting:

Why Planning Matters:

Without explicit planning instructions, LLMs tend to jump straight into execution. This works for simple tasks but fails for complex ones. Planning forces the LLM to think about the entire task before starting, identify dependencies, and allocate effort appropriately.

Like how a good architect does not start laying bricks on day one. They first survey the land, draw blueprints, calculate materials, and plan the construction sequence.

Planning Prompt Patterns:

• Task Decomposition: "Before solving, break this into numbered sub-tasks. For each sub-task, specify what input it needs and what output it produces."
• Dependency Mapping: "Identify which sub-tasks depend on others. Which can run in parallel?"
• Resource Estimation: "For each sub-task, estimate the complexity and what tools or information you will need."
• Milestone Setting: "Define checkpoints where you will verify progress before continuing."

Example - Wedding Planning Agent:

User: "Plan a budget wedding in Delhi for 200 guests under 10 lakhs"

Agent Plan:
1. Budget Allocation (no dependencies)
   - Venue: 3L, Food: 3L, Decor: 1.5L, Photography: 1L, Misc: 1.5L
2. Venue Research (needs: budget from step 1)
   - Search venues in Delhi within 3L budget
3. Vendor Shortlisting (needs: venue from step 2)
   - Caterers, decorators, photographers near venue
4. Menu Planning (needs: budget, guest count)
   - 200 guests x Rs 1,500 per plate
5. Timeline Creation (needs: all above)
   - Week-by-week checklist
6. Review and Optimize (needs: complete plan)
   - Check total against budget, adjust if needed

What is Reflection?

Reflection is instructing the LLM to evaluate its own output, identify issues, and improve before presenting the final answer. It is like telling a student "Write your essay, then re-read it and fix any mistakes before submitting."

Without reflection, LLMs are like students who submit their first draft. With reflection, they revise and produce significantly better work.

Reflection Patterns:

• Self-Critique: "Review your answer. What are 3 potential issues or inaccuracies? Fix them."
• Rubric-Based: "Grade your answer on: accuracy (1-5), completeness (1-5), clarity (1-5). If any score is below 4, improve it."
• Devil's Advocate: "Argue against your own answer. What would someone who disagrees say? Now address those counterarguments."
• Perspective Shift: "Re-read your answer as if you are the target audience. Is it clear? Is anything confusing?"

Reflexion Framework (Research Paper):

Reflexion adds a memory component to reflection. After each attempt, the agent writes a reflection note (what went wrong, what to do differently). On the next attempt, it reads its own past reflections to avoid repeating mistakes.

Like a cricketer watching video replays of their dismissals. Each time they bat, they remember past mistakes and adjust their technique.

When Reflection Helps vs Hurts:

Tool Use in Agentic Prompting:

The most powerful agentic capability is knowing when to stop relying on internal knowledge and reach for an external tool instead. This prevents hallucination and grounds the agent in real data.

Like a doctor who knows when to order a blood test instead of guessing the diagnosis. The skill is not in using the tool - it is in knowing WHEN to use it.

Teaching Tool Awareness Through Prompting:

• Capability Boundary: "If you are not sure about current data, prices, or recent events, use the search tool. Do not guess."
• Tool Selection Logic: "For math calculations, use the calculator. For code execution, use the code runner. For factual queries, use web search."
• Confidence Threshold: "If your confidence in the answer is below 80%, use a tool to verify before responding."
• Parallel Tool Use: "If multiple independent pieces of information are needed, request all tool calls at once."

Tool Use Decision Tree:

Question arrives ->
  Is this about current/real-time data?
    YES -> Use search/API tool
    NO -> Is this a calculation?
      YES -> Use calculator/code tool
      NO -> Am I confident in my knowledge?
        YES (90%+) -> Answer directly
        NO -> Use search to verify
        PARTIALLY -> Answer + flag uncertainty

Common Tool Categories:

The Complete Agentic System Prompt Structure:

• Role: Define who the agent is and its expertise area
• Planning Instructions: How to break down tasks, when to plan vs act immediately
• Tool Descriptions: Available tools with clear usage criteria
• Reflection Rules: When to self-check, quality criteria, max reflection rounds
• Output Format: How to structure the final response
• Guardrails: What the agent must NOT do, safety boundaries

Example - E-Commerce Research Agent Prompt:

You are a product research agent for Indian e-commerce.

PLANNING:
- For any product comparison, first list all products to compare
- Identify what attributes to compare (price, specs, reviews)
- Plan which tools to call for each product

TOOLS:
- search(query): Search the web for product info
- price_check(product, platform): Get current price
- reviews(product): Get aggregated review scores

REFLECTION:
- After gathering all data, verify completeness
- Check if any critical attribute is missing
- Ensure prices are current (not cached/old)

OUTPUT:
- Comparison table with all attributes
- Clear recommendation with reasoning
- Mention any caveats or uncertainties

Inner Monologue Pattern:

One powerful technique is the inner monologue - instructing the agent to think out loud in a structured format before acting. This makes the agent's reasoning transparent and debuggable.

For each step, think in this format:
THINKING: What do I need to do next and why?
DECISION: Should I use a tool or respond directly?
ACTION: [tool call or direct response]
OBSERVATION: What did I learn from this?
REFLECTION: Is the quality good enough? Any issues?

Q: What are the three pillars of agentic prompting?

Q: How does Reflexion differ from simple self-critique?

Q: When should an agent use a tool vs answer from its own knowledge?

Q: What is the Inner Monologue pattern in agentic prompting?

Q: How many reflection rounds should an agent do?