# Amdahl's Law

> The serial part of a system sets the ceiling on the speedup available from parallel work.

- HTML version: https://robbiepalmer.me/ideas/amdahls-law
- Source: https://en.wikipedia.org/wiki/Amdahl%27s_law

Amdahl's Law puts a hard limit on gains from parallelism. If a fraction `s` of a task must remain
serial, the maximum speedup with `N` workers is:

`1 / (s + (1 - s) / N)`

As `N` grows, the parallel portion gets cheap and the serial portion dominates. Additional workers
add coordination without moving the finish line much.

The law comes from parallel computing, but it is a useful model for teams and coding agents. Ten
agents can produce changes at once. If one person must specify, review, integrate, and release all
of them, those steps become the real system. Faster generation makes the bottleneck more visible.

## Questions it prompts

* Which part of this workflow cannot run concurrently?
* Can we remove that step, split its ownership, or make its evidence cheaper to review?
* Are we adding parallel capacity after the useful speedup has already flattened?

---

Markdown index of this site: https://robbiepalmer.me/llms.txt