Law 3 of 12 in the Hacker Laws for Agentic Software Engineering series — read the index. Previous: Brooks' Law · Next: Gall's Law.
The Law
Amdahl's Law is a formula which shows the potential speedup of a computational task which can be achieved by increasing the resources of a system. Normally used in parallel computing, it can predict the actual benefit of increasing the number of processors, which is limited by the parallelisability of the program. (hacker-laws)
The Key Insight for Agentic Software Engineering
Agent swarms are parallel processors, and Amdahl's Law applies to them exactly as it applied to CPUs: the speedup from adding agents is bounded by the fraction of the task that can actually be parallelised, and even a task that is 95% parallelisable caps out long before the swarm grows large. In an agentic pipeline, the parallelisable fraction is the independent work — the feature edits, the searches, the test runs that can be handed to separate workers. The serial fraction is everything that must happen in one context: the plan, the merge, the context building, and — most importantly — the verification.
The verifier is usually the serial fraction, and it is the one this blog keeps proving is the bottleneck: Fowler's retreat made it the headline — "code generation is no longer the bottleneck — verification is" (Verification Is the Bottleneck). Spawn a thousand agents to write code faster and the pipeline still drains through the single verifier that has to check it all — the Sonar AC/DC finding is the same shape: verification is where the 3-5x velocity boost rots. This is why the evaluator with hands is reserved for the final gate: it is the most valuable serial step, and running it in parallel with itself is the one thing Amdahl's Law says you cannot do.
The ASE reading of Amdahl's Law: measure the serial fraction of your agent pipeline — the verifier is usually it, and no number of workers beats the ceiling it sets. The harness-level lever is to raise the parallel fraction, not to add workers: decomposition turns one serial task into many parallel ones, mocks let independent runs proceed without waiting on shared infrastructure, and verification must be layered so the cheap checks run in parallel and only the expensive judgment is serial.