Day 8 – Planning In Agents (re Act, Plan-and-execute) | @swati_goyal_911 | QuizMaker

Series or course

Agentic AI

Why Planning Is the Real Brain of an Agent 🧠

Here’s a hard truth:

An agent without planning is just a fancy chatbot with tools.

What separates:

• ❌ brittle tool-calling bots

• ✅ reliable, autonomous agents

…is planning.

Planning answers three critical questions:

1. What should I do first?

2. What should I do next?

3. When should I stop?

Without planning, agents:

• call tools randomly

• repeat steps endlessly

• hallucinate progress

• burn tokens and money 🔥

Two Dominant Planning Paradigms (2026)

In production systems today, 95% of agent planners fall into one of these models:

Let’s unpack both — deeply.

1️⃣ ReAct: Reasoning + Acting in a Loop 🔁

Mental Model:

The agent thinks out loud, takes an action, observes the result, then thinks again.

ReAct Loop Diagram

[Reason]
   ↓
[Action → Tool Call]
   ↓
[Observation]
   ↓
[Reason]
   ↓
[Next Action]

This loop continues until a stop condition is met.

Example: Research Agent

Task: “Find the latest AI regulations in the EU and summarize key risks.”

ReAct-style flow:

1. Reason: I need authoritative sources → search web

2. Act: Call search tool

3. Observe: Found EU AI Act summaries

4. Reason: Need official language → open regulation text

5. Act: Fetch document

6. Observe: Long legal text

7. Reason: Extract risk sections

8. Act: Summarize

Why ReAct Works So Well

✅ Adapts to new information
✅ Handles unknown paths
✅ Feels intelligent and flexible

Where ReAct Fails ⚠️

CTO Insight:

ReAct is powerful — but dangerous without guardrails.

2️⃣ Plan-and-Execute: Think First, Act Later 📋

Mental Model:

Create a plan, then follow it step-by-step.

Plan-and-Execute Diagram

[Create Plan]
   ↓
[Step 1]
   ↓
[Step 2]
   ↓
[Step 3]
   ↓
[Done]

The agent commits to a roadmap before acting.

Example: Data Pipeline Agent

Task: “Generate a weekly sales report.”

Generated Plan:

1. Fetch sales data from database

2. Clean missing values

3. Aggregate by region

4. Generate charts

5. Write summary

Execution follows this sequence strictly.

Why Enterprises Love This Model

✅ Predictable execution
✅ Easier debugging
✅ Cost and time bounded
✅ Easier compliance

Where It Breaks ⚠️

CTO Insight:

Plan-and-Execute trades adaptability for reliability.

ReAct vs Plan-and-Execute — Head-to-Head

The Real-World Pattern: Hybrid Planning 🧩

Modern agents combine both approaches.

Hybrid Architecture

[High-Level Plan]
   ↓
For each step:
   └─ ReAct loop inside

Example: Customer Support Agent

High-level plan:

1. Understand issue

2. Check account

3. Diagnose problem

4. Propose solution

Inside “Diagnose problem”, the agent uses ReAct to explore logs, FAQs, and tools dynamically.

This is the sweet spot.

Planning Anti-Patterns to Avoid 🚫

❌ Letting agents plan endlessly
❌ No maximum step limits
❌ No success criteria
❌ No fallback paths
❌ Confusing reasoning with action

If you see these — your agent will fail at scale.

A Simple Planning Checklist ✅

Before deploying an agent, ask:

• What is the goal state?

• What defines done?

• How many steps max?

• What happens on failure?

• Can humans interrupt?

If you can’t answer these — stop.

Final CTO Takeaway

Planning is not an implementation detail.

It is the architecture of intelligence.

Choose:

• ReAct when the path is unknown

• Plan-and-Execute when reliability matters

• Hybrid when building real products

Agents don’t fail because models are weak.

They fail because planning was an afterthought.