The 'Planning' Step: Cost vs. Performance

In agentic AI, the Planning Step is where the model looks at a goal and says: "To solve this, I must do X, then Y, then Z."

This is the most "Intelligent" part of the loop, but it is also the most Token-Hungry. If an agent re-plans after every single tool call, you are paying for a "Deep Reasoning" turn 10 times. If you don't plan enough, the agent gets lost and wanders into expensive error loops.

In this lesson, we learn the two types of planning—Static vs. Dynamic—and how to choose the right one for your token budget.

---

1. Static Planning (The 'Batch' Approach)

In a static plan, the agent generates a full sequence of steps at the Beginning and then executes them blindly.

• Pro: Low token cost. Only one planning turn.
• Con: Brittle. If Step 1 fails, Step 2 and 3 might be impossible, leading to wasted tokens on "Invalid Actions."
• Best For: Highly predictable workflows (e.g. "Extract data, format it, save to DB").

---

2. Dynamic Planning (The 'Reactive' Approach)

In a dynamic plan, the agent re-evaluates its strategy after every "Observation."

• Pro: High accuracy. Flexible in the face of errors.
• Con: High token cost. Every turn is a full "Reasoning" turn.
• Best For: Open-ended research, coding assistants, and complex customer support.

graph TD
    subgraph "Dynamic Planning"
        P1[Plan 1] --> O1[Action 1]
        O1 --> P2[Plan 2 - Revised]
        P2 --> O2[Action 2]
    end
    
    subgraph "Static Planning"
        SP[Plan A, B, C] --> SO1[Action A]
        SO1 --> SO2[Action B]
        SO2 --> SO3[Action C]
    end

---

3. The "Hybrid Plan" Strategy (The Sweet Spot)

The most token-efficient architecture uses Conditional Planning Points. Instead of re-planning every turn, only trigger a planning turn if:

1. A tool returns an Error.
2. A tool returns Zero Results.
3. A specific Milestone is reached.

This reduces planning tokens by 60-80% while maintaining 95% of the accuracy of a fully dynamic agent.

---

4. Implementation: The Milestone-Based Planner (Python)

Python Code: The Thrifty Planner

def agent_executor(goal):
    # Pass 1: Initial Plan (Static)
    plan = call_llm(f"Goal: {goal}. Output: List of Steps.")
    
    for step in plan:
        result = execute_step(step)
        
        # WE ONLY RE-PLAN ON FAILURE
        if "ERROR" in result:
            print("Deviation detected! Re-planning...")
            # This is where we spend the big reasoning tokens
            plan = call_llm(f"Status: {result}. New Plan:") 
            continue
            
        print(f"Step {step} complete. Moving on.")

---

5. Token Savings in Multi-Step Workflows

By using Static Planning for the "Happy Path," you save tokens on the vast majority of your users. You only pay for "Extra Intelligence" when the environment proves difficult.

Think of it as "Level of Detail" (LOD) for AI.

• Level 1: Follow the script (Cheap).
• Level 2: Re-evaluate if confused (Expensive).

---

6. Planning and Context Window Pressure

Planning generation is "Output Heavy." As we know from Module 1, output tokens are expensive. By constraining the planning turn to only output JSON steps (rather than a 500-word essay about the plan), you directly reduce the $ costs of the most frequent turn in your agent's life.

---

7. Summary and Key Takeaways

1. Static is Baseline: Start with a fixed sequence of steps to save tokens.
2. Dynamic is insurance: Use re-planning only when the environment changes or an error occurs.
3. Control Output: Use structured formats for plans (JSON/Lists) to prevent planning verbosity.
4. Budgeted Reasoning: Give your planner a "Max Reasoning Length" constraint.

In the next lesson, Throttling and Budgeting for Autonomous Agents, we conclude Module 9 by learning how to put a "Hard Cap" on agentic spending.

---

Exercise: The Re-planning Audit

1. Run an agent that has to find information from 3 different websites.
2. Version A: Re-plan after every single website search.
3. Version B: Use a static list of 3 websites to search.
4. Compare the Total Token Bill.

• If the websites were correctly found in both versions, how much did you "Overpay" for the intelligence in Version A?
• (Usually, Version A is 3x more expensive for the exact same result).

---

Congratulations on completing Module 9 Lesson 4! You are now a strategic agent architect.