Andrej Karpathy described a simple pattern this year: point an agent at a folder of markdown, let it read raw sources, and let it write and rewrite a wiki that summarizes what it learns. It caught on fast, because it works.
wiki/transformer-throughput.md
The wiki page summarizes a source on inference optimization.
The source says the approach doubles throughput. The source reports a 40 percent throughput gain, not double.
An agent edit, caught in review. This is the whole product.
In Karpathy's design, the agent owns the wiki layer end to end. It ingests a source, writes a page, updates an index, and moves on. Nothing forces it to check its own work.
That is fine when the wiki is a scratchpad only you look at. It stops being fine the moment a wiki page is wrong and something downstream, another agent, a teammate, a decision, treats it as true. A knowledge base nobody reviews is a knowledge base nobody actually trusts.
The gap is not the pattern. The pattern is good. The gap is that "agent owns the wiki" has no step where a human looks at what changed before it counts as settled.
The source of the pattern: sources are immutable, the wiki is agent-written, and an index and a log keep it organized.
datasciencedojo.comA step-by-step guide to building an LLM wiki from scratch.
academy.dair.aiAn explanation of why the pattern works as a knowledge base for agents.
github.comAn open-source implementation connecting Claude to an Obsidian vault.
kennethreitz.orgAn essay on running Claude Code against an Obsidian vault as working memory.
Every one of these setups has the same shape: an agent with write access to the one place your team's knowledge accumulates. The wiki can drift from the sources it is supposed to summarize. The index can go stale. A page can quietly get worse with every edit, and nobody would know until it mattered.
The fix people reach for is usually "check it later," which in practice means never. Reviewing an agent's wiki edit today means opening a git diff and reading markup, not the page it produces. Most people skip that step, so the review never happens at all.
Open the wiki folder in Stet instead of a plain editor. When your agent ingests a source and writes a page, the edit shows up as a proposal you accept or reject, read in plain English, not a diff.
Leave an AI note on a page telling the agent what to fix next time it passes through. And because every proposal is saved, the document's history shows you exactly how a page got to its current state, source by source.
Download the starter kit. It has a sample source, an empty wiki, and a README that walks you through your first review.
sources/example-source.md and ask it to ingest the source into the wiki.