Agents Drift When Context Is Rebuilt

Humans carry context between slices. When an engineer finishes the first slice of a feature and moves to the second, they bring the problem, the prior choices, the constraints, and the style of the code with them. Incremental development works because the human is the continuity layer.

Agents are not that continuity layer. Every time an agent starts a new slice, it rebuilds context from whatever artifacts it can read: the current code, the spec, recent history. That rebuild is never complete and never neutral. It introduces small reinterpretations, style drift, and quiet assumptions that accumulate across slices.

Picture a novel written chapter by chapter by an author who wakes up each morning with no memory of the previous chapters, handed only the manuscript-so-far to read. Every chapter is competent. But the protagonist's name shifts from "Marco" to "Mark" by chapter four, the tone moves from literary to pulpy, and a subplot introduced in chapter two is quietly forgotten by chapter six. The novel is unreadable as a whole even though every chapter, on its own, is well written.

Agent-built codebases drift the same way. Naming conventions shift, abstractions diverge, assumptions made on slice one get quietly contradicted on slice five. Each diff looks clean in review. The whole artifact looks like it was written by a committee of amnesiacs. The drift is usually invisible slice by slice. It shows up when you step back and look at the complete result.