Patterns the agent has seen ten million times produce confident, correct edits. Bespoke abstractions and metaprogramming layers — efficient for humans — push the agent outside its statistical confidence interval.

Example: A vanilla repository pattern accepts an agent's edit on the first try. A clever code-generation framework requires four iterations and a senior reviewer to undo the second one.

A comprehensive executable test suite is no longer a quality artifact. It is the verification mechanism by which the organization decides whether to accept the agent's work.

Example: A repository with broad behavioral coverage merges agent diffs in minutes. A repository without it routes every diff through senior review, and the agent's productivity becomes the reviewer's queue.

Cross-boundary dependencies multiply the inference complexity an agent must hold simultaneously. Defect rate climbs with every additional module the agent must reason across.

Example: An edit confined to a single bounded module ships clean. The same edit cutting across three modules with shared state produces a confident output that compiles, passes tests, and corrupts state in production.

The hidden cost of rework, debugging, and rollback when agents encounter unfamiliar patterns is real and measurable. Naming the tax is the first step to allocating against it.

Example: Two repositories serve the same business function. One produces a clean adoption curve. The other absorbs senior labor at a rate that nobody costs because nobody measures it.

A control regime designed around human-reviewed human-authored change does not cover the failure mode of confident, plausible, semantically incorrect agent output. The regime must be re-specified.

Example: A traceability standard built around named human commits cannot answer the question of which agent produced which line under which prompt — and the auditor's question is exactly that.