The end of "Just ask Sarah"

Every engineering team has someone like Sarah. She knows why the service was split that way, why the obvious fix was rejected, why a weird constraint shaped the current design. Humans can walk over and ask her. AI agents can't. Simon Aronsson makes this case in a recent essay: as agents start writing and modifying code, the cost of undocumented decisions goes from recoverable to permanent.

When an agent opens a task and scans the codebase, it looks for patterns. If there's no Architecture Decision Record explaining why services are structured a certain way, no spec describing intent, the agent just finds the dominant pattern and extends it. And why wouldn't it? The code is there, so a decision was made. The reasoning just wasn't written down. The agent inherits the output without the reasoning and fills in the blanks. It might keep reproducing a design pattern born from a scaling constraint that disappeared two years ago. As noted in The Invisible Blast Radius, codebases often lack the deterministic structure needed for reliable AI output.

Aronsson argues that documentation needs to shift from courtesy to execution context. ADRs, specs, and playbooks introduce friction that stops agents from blindly extending patterns. Without them, you risk what one Hacker News commenter described as agents hallucinating ADRs for legacy code, rationalizing technical debt as intentional architecture. Future developers and AI systems then treat these invented records as ground truth.

The problem isn't new. Organizations have always undervalued people who hold institutional knowledge. One commenter described being a "Sarah" as becoming their company's first agent, given endless work with just enough authority for tasks but not enough respect to address root causes. What changes with AI is timing. A human engineer accumulates context over time and carries it between sessions. Agents start from zero every time. If the organization hasn't produced anything worth retrieving, a vector database full of stale tickets won't save you—unlike a tool like GBrain that compounds knowledge over time.