Compare
Symptom
Human

A type — class or interface — with too many fields and methods; multiple unrelated responsibilities under one name.

Agent

A type definition — class or interface — with so many fields and methods that the agent cannot load it as a coherent unit; multiple unrelated responsibilities sit under one name.

Goal
Human

Each type has one cohesive purpose; methods cluster around fields they actually use.

Agent

Each type has one cohesive purpose; the agent loads a small focused file to reason about any single behavior.

Smellier version

class Order {
// lineItems, totals, customer info, shipping address, audit log, ...
}

Fresher version

class Order { /* lineItems, totals */ }
class Customer { /* name, email */ }
class Shipping { /* address, track */ }
Example source: Illustrative example written for this site, not a quotation from any source.
Pressure
Human

Cognitive load: every reader pays for fields they don't care about; merge conflicts spike; testing is unfocused.

Agent

Cognitive context inflates with every irrelevant member; the agent reading any single method must skim past unrelated fields and helpers to find what it needs.

Tradeoff
Human

Splitting a large type creates collaborator relationships that the team must learn; some types are large because they encode a genuinely large domain concept that resists decomposition.

Agent

Splitting creates new collaborator relationships the agent must trace across files; what was one load now becomes N loads and the agent must understand how the pieces interact.

Relief
Human

Smaller, more testable units; clearer ownership; faster reading.

Agent

Smaller, focused units; the agent's reasoning context per method is tight; tests target one concern at a time.

Trap
Human

Splitting every large class on superficial groupings (methods that happen to share a prefix, fields that happen to be related) without honoring real cohesion boundaries — produces fragments that don't make sense apart.

Agent

Splitting on superficial groupings produces fragments that don't cohere on their own — the agent now navigates several files to understand what was previously one coherent (if large) concept.