The Planner-Worker Pattern
Everyone Gets Wrong

Common Mistakes & How to Fix Them
Most implementations fail because they treat it as sequential, not iterative.

Wrong vs Right Approach

❌ What Most Do
Planner: Create full plan
↓
Worker 1: Execute step 1
↓
Worker 2: Execute step 2
↓
Worker 3: Execute step 3
↓
Done
✅ What You Should Do
Planner: Next step only
↓
Worker 1: Execute & report
↓
Planner: Re-evaluate
↓
Worker 2: Execute & report
↓
Planner: Adjust

Common Mistakes

Mistake #1
Planning Everything Upfront
Creating a complete plan before execution. Reality changes, plan becomes obsolete.
Mistake #2
No Feedback Loop
Workers execute blindly without reporting back. Planner can't adjust.
Mistake #3
Treating Failures as Fatal
One worker fails → entire system fails. No retry, no alternative paths.

State Management

What to Track
Current goal
Steps completed
Results so far
Failed attempts
How to Track
Shared state object
Update after each step
Pass to next worker
Planner reads state
What Workers Return
Success/failure status
Result data
Error details (if failed)
Recommendations
What Planner Does
Read worker results
Update understanding
Decide next action
Adjust if needed

When Workers Fail

Strategy 1: Retry with Backoff
Worker fails → Wait → Retry with same input
Best for: Transient errors (network, rate limits)
Strategy 2: Alternative Approach
Worker fails → Planner tries different worker or method
Best for: Tool-specific failures (API down, access denied)
Strategy 3: Simplify Task
Worker fails → Planner breaks task into smaller pieces
Best for: Task too complex (context overflow, ambiguity)
Strategy 4: Human Escalation
Multiple failures → Ask human for help or clarification
Best for: Unrecoverable errors, ambiguous requirements

Weekend Project

Build It This Weekend
Goal: Planner orchestrates 3 workers to complete a research task
Planner
Decomposes task, decides which worker to call, re-evaluates after each step
Worker 1: Data Retrieval
Makes API calls, returns raw data or error
Worker 2: Data Processing
Transforms data, performs computation, returns results
Worker 3: Summary
Takes processed data, generates natural language summary

Key Principles

1. Plan one step at a time, not the entire journey
2. Workers report back, planner re-evaluates
3. Failures are opportunities to adjust, not endpoints
4. State flows through the system, not just data
Iterative, not sequential. That's the difference between brittle and robust.
💡
Implementation Examples
community.nachiketh.in
🎓
Production Patterns
bootcamp.nachiketh.in