Planning & Reasoning

Problem Decomposition

Break complex goals into simpler sub-goals:

Goal: "Organize a conference"

Decomposition:
1. Venue
   1.1 Find venue (capacity 500)
   1.2 Book venue
   1.3 Confirm catering
2. Schedule
   2.1 Plan agenda
   2.2 Invite speakers
   2.3 Send calendar invites
3. Marketing
   3.1 Create website
   3.2 Social media campaign
   3.3 Email newsletter

Agent executes sub-goals in dependency order!

Planning Strategies

Linear Plan

Task 1 → Task 2 → Task 3 → Goal

Simple, but brittle (fails if any step fails)

Conditional Plan

Task 1 → Check result
  If success → Task 2
  If failure → Task 1_retry or Alternative
  
More flexible, handles failures

Hierarchical Plan

Abstract goal
    ↓
Detailed sub-goals
    ↓
Even finer sub-tasks
    ↓
Executable actions

Example: "Write essay"
  → Research topic (find 5 sources)
    → Search articles
    → Download PDFs
  → Create outline
  → Write sections

Reasoning Under Uncertainty

Agents deal with incomplete information:

Agent has options to search, but uncertainty about which is best:

Option A: Book flight X (cheaper, but no direct route)
Option B: Book flight Y (more expensive, direct route)

Agent reasoning:
- Check user's budget (allows both)
- Check user's time constraints (prefers direct)
- Recommend Option Y

Agents reason about tradeoffs!

Handling Failures

What happens when agent fails?

Plan: [Book flight]
  Step 1: Search flights ✓
  Step 2: Check availability ✗ (flight full)
  
Response:
  Option 1: Try alternative dates
  Option 2: Try alternative airlines
  Option 3: Ask user for flexibility
  
Adaptive agents recover from failure!

Self-Reflection

Agent reviews its own reasoning:

Step 1: Attempt goal with approach A
Step 2: Observe failure
Step 3: Reflection: "Approach A failed because X. I should try approach B instead."
Step 4: Execute approach B
Step 5: Success!

Reflexion agents learn and improve!