← Patterns

Amdahl's Law

Core Concept

“The speedup of a program using multiple processors is limited by the fraction of the program that must remain sequential.”

You can parallelize all you want, but the sequential bottleneck determines your maximum speedup. If 50% of the work is sequential, infinite parallelism only gives you a 2x speedup. Optimize what’s actually slow, not what’s easy to optimize.

Where It Applies

Pipeline Optimization

If the bottleneck is human review (sequential), making agent stages faster (parallel) yields diminishing returns. Improvements that reduce human friction — notifications, better formatting, executive summaries — may deliver more wall-clock speedup than any agent optimization.

Team Scaling

Throwing more engineers at a project hits Amdahl’s Law fast. The sequential parts — architecture decisions, code review, deployment coordination — don’t speed up with more people. This is why Brooks’s Law (“Adding manpower to a late software project makes it later”) is a corollary of Amdahl’s Law applied to human systems.

Process Improvement

Before optimizing any step, ask: “Is this step on the critical path?” If not, making it faster doesn’t make anything faster.

The “Electricity or Blood?” Connection

Amdahl’s Law tells you where to apply electricity. Automate the bottleneck, not the thing that’s easiest to automate.

The Test

Before optimizing a stage, ask: “Is this stage on the critical path, or is a human checkpoint the bottleneck?” Optimize what’s actually slow.

Cross-References

• Software-Laws — hub for all seven laws and how they interact
• Parallel-Projects-Sequential-Tasks — Dave’s corollary: coordination costs are superlinear, making parallelism net negative past a crossover point (extends Amdahl from “limited speedup” to “negative speedup”)
• Goodhart’s Law — tension: removing quality checks speeds things up (Amdahl) but exposes you to metric gaming (Goodhart)