amdahls-law ADG-0001Amdahl's law
“The speed-up from parallelising a task is bounded by the fraction of it that must stay serial.”
Origin
Gene Amdahl, 1967, on multiprocessor computing. A formula rather than an observation: with fraction p parallelisable, speed-up is at most 1/(1-p).
Why it should, or should not, apply to models
Multi-agent systems are pitched as parallelism, and the law says where the ceiling is: whatever cannot be split — integration of the parts, the verification pass, the human review, the one shared piece of state — bounds the gain no matter how many agents run. The prediction is that as agents get cheaper the serial fraction (usually verification) dominates wall-clock and cost, and that a system's throughput is better improved by shrinking that fraction than by adding agents. It is also a warning about the catalog's own ambition: an evidence pipeline is bounded by the serial step of a person reviewing.
Standing: Untested
Standing counts claims a person has reviewed. Claims reviewed by AI alone are listed and marked, and do not move it.
No evidence filed either way.
Searched 2026-09-12: Nothing in the catalog measures the serial fraction of a multi-agent workflow or plots gain against agent count. The nearest claims are about long-horizon single agents, not parallel ones. Nearest misses: long-horizon-tool-calling-agents-tasks-needing-five-or-more-calls.
A break is worth more than a hold. If you have seen this fail for a model, with the setup written down, that is the most useful thing you can add here.