lehmans-law-of-increasing-complexity ADG-0017

Lehman's law of increasing complexity

“As a system evolves its complexity increases unless work is done to maintain or reduce it.”

Also: Manny Lehman's law

Origin

Meir "Manny" Lehman, 1974 onward, the second of his laws of software evolution, from studying IBM's OS/360. About long-lived software in real organisations, where each change adds structure and nobody removes any.

Why it should, or should not, apply to models

An agent's context is a small evolving system: guide files, accumulated tool output, notes, prior turns. Each run adds to it and nothing removes anything by default. The law transfers if that accumulation has a cost, and the catalog says it does — irrelevant tokens degrade reasoning on their own — so the "work done to reduce it" clause becomes a design requirement: context has to be pruned, paged or compiled, or performance decays. The twist is that the maintenance work itself needs capability; handing a weak model the pruning tools makes it worse.

Standing: Holds

Standing counts claims a person has reviewed. Claims reviewed by AI alone are listed and marked, and do not move it.

Where it holds only under a condition

Claims that keep the heuristic but bound it.

Handing an agent context-editing tools (delete, summarize, note, fold) by prompt alone makes mid-size models worse than leaving their context untouched — Qwen3-8B, Qwen3-14B and Gemma4-E4B all scored lower on long-context QA with the tools than without them; only after fine-tuning on trajectories that use the tools does the toolset become a gain.
Keeping its own context cleanReviewed by AI
The maintenance work is not free: mid-size models given pruning tools by prompt alone got worse. The law holds; the fix needs capability or training.

Where it holds

Claims showing the heuristic applying to models.

Adding irrelevant tokens alone degrades reasoning, which is the mechanism by which a bloated context hurts an agent.
Keeping its own context cleanReviewed by AI and Russ Miller
The cost side of the law: accumulated irrelevant context degrades reasoning by itself, so growth without maintenance has a price.
Managing context explicitly, paging information in and out of a bounded window, sustains performance on tasks that exceed the window.
Keeping its own context cleanReviewed by AI and Russ Miller
The maintenance side: explicitly paging context in and out sustains performance where unmanaged growth would not.

Add a case where it holds or breaks

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.