braesss-paradox ADG-0005

Braess's paradox

“Adding capacity to a network can make its overall performance worse.”

Origin

Dietrich Braess, 1968, on road networks: a new road can raise everyone's journey time because each driver's selfish choice shifts the equilibrium. About systems where local improvements change global behaviour.

Why it should, or should not, apply to models

The paradox transfers wherever adding a resource changes what the model attends to rather than simply adding an option. More tools in the prompt means more schema to read and more wrong tools to pick; more agents with a shared channel means more routes for an error to spread; more context-management tools means more ways to damage the context. The mechanism is not selfish drivers but a fixed attention budget and a model that uses what it is given, so the prediction is that capacity added without a corresponding reason to ignore it will often cost more than it returns.

Standing: Untested

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

Where it holds

Claims showing the heuristic applying to models.

When an agent must work over a very large tool catalogue, wrapping each tool in a natural-language interface that resolves the API schema internally — so raw schemas are retrieved on demand rather than enumerated in the model's prompt — contributes more to multi-step tool- call success than the surrounding planner, router, or verifier stages.
Using the tools it is givenReviewed by AI
Enumerating a large tool catalogue in the prompt hurt; hiding the schemas behind a natural-language wrapper — removing capacity from the model's view — helped more than any other stage.
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
Adding context-editing tools by prompt alone made mid-size models worse than having no such tools.
In multi-agent LLM systems where one agent starts with an erroneous shared belief, broadcasting every agent's full context at every step raises the rate of false assertions above doing no synchronization at all, because the error propagates to agents that were previously correct — and the harm appears only in tasks where one wrong fact cascades across semantically linked dimensions (destination to airport to weather), not where agent contexts are largely orthogonal.
Stating false facts confidentlyReviewed by AI
Adding communication links (full broadcast) raised the false-assertion rate above no links at all.

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.