Code Review Is Dead: Agents Write Faster Than Humans Read

The teams covered by The Pragmatic Engineer are adopting risk-based triage: critical paths (payments, authorization, and schema migrations) are reviewed manually, while everything else goes through an AI reviewer or is not read line by line at all. Instead of the diff, they check the outcome: whether tests pass, the data schema remains intact, and system behavior has not changed where it is measured. Control shifts from how code is written to the result—where it can actually be verified automatically.

Dex Horthy of HumanLayer describes the root cause differently: an agent does not become ineffective on its own; it degrades when its context window is cluttered with irrelevant fragments—he calls this trajectory poisoning. In his assessment, managed context makes development agents 2–3 times more productive without sacrificing quality; unmanaged context produces code that no one can later review because it was written from a mix of outdated instructions and random files the agent picked up along the way.

Context engineering is not an option for enthusiasts; it is what makes risk-based review meaningful: when the agent's input is clean, its output is predictable and easier to validate with metrics rather than line by line.

The result business sees—“AI completed the migration in two weeks”—looks simple precisely because it rests on engineering no one sees: clean agent input context, test coverage that catches regressions automatically, and gates that prevent code from advancing until it passes validation.

At KT.Team, we build precisely this invisible layer: LLM & Security Gateway to control what the agent sees and where it sends data; MCP for structured access to internal systems instead of arbitrary chat context; and RAG to add up-to-date documentation instead of outdated prompt instructions. Each component reduces the chance of trajectory poisoning at the input—and therefore reduces the amount of output code that requires manual review.

Line-by-line review of AI-generated code applies an industrial-era process to a pipeline that produces an order of magnitude more. Companies that continue either slow agents to human speed and lose the point of using them, or merge unchecked code and face incidents. The practical alternative is risk-based triage, outcome-based validation (tests, schemas, and metrics) instead of diff review, and investment in clean agent input context—not just faster responses.

Anyone who does not adapt the review process to the new speed of code production is postponing a bill that will arrive with the first serious incident.