Documentation Drift Is Now a Runtime Bug for AI Coding Agents

Documentation drift used to be a nuisance for human developers — a day wasted chasing down why the billing path was moved. Now it's a runtime bug. AI coding agents read AGENTS.md, CLAUDE.md, wikis, and runbooks as operating instructions. A stale architecture note doesn't just confuse a junior dev; it gets serialized into a pull request that looks productive but is wrong.

LangChain recently released OpenWiki, an open-source agent and CLI that generates and maintains repo documentation specifically for coding agents. The tool creates an openwiki/ directory, appends agent instructions, and supports openwiki --init and openwiki --update with a GitHub Action template. It's a direct acknowledgment that repo docs are now part of the execution path, not just onboarding material.

Docs Are Part of the Runtime Now

The hot file — the prompt cache or AGENTS.md — should contain only rules that apply to every run: build commands, service boundaries, security constraints. It should not become a dumping ground for every architecture decision ever made. Chroma's context-rot research shows that LLM performance degrades as context grows across 18 tested models. More context isn't better; it's noisier and more likely to latch onto stale paragraphs.

The Codified Context paper (arxiv 2602.20478v1) demonstrates a 108,000-line C# system with a three-tier context architecture: hot-memory constitution, specialized agents, and a cold-memory knowledge base. The paper treats agent-facing documentation as load-bearing infrastructure — it should have owners, change in review, be easy to diff, and have a way to fail.

Drift Is the Failure Mode

A stale runbook causes an incident assistant to work against the wrong dashboard. An agent runs tests the team no longer has on disk and pushes incorrect code. The fix is to put agent-facing docs in the same loop as agent evaluation. LangChain's Pendo case study shows Novus achieving a 90%+ success rate on PM-reviewed evals by connecting product analytics, session replay, and LangSmith traces to code fixes. If a reviewer repeatedly corrects the same mistake, that knowledge belongs in the agent-facing documentation.

Mixpanel reported 50% more pull requests with the same engineering team after AI entered the workflow. Faster code makes stale context more expensive. Every pull request now requires documentation updates for agents, and every wrong premise generates rework loops. The whole system sounds busy while being wrong.

Treat repo docs like config files: versioned, reviewed, and owned. The alternative is an AI that confidently ships the wrong change.