Arriving knowing the room
Before LSP, every code editor built its own understanding of every language. M editors times N languages meant M×N integrations. LSP collapsed that — one protocol, and any editor could work with any language.
Agents have the same problem with knowledge. Every agent session starts from scratch, guessing at what’s in your workspace. Grep works for exact strings. Anything thematic — “what have I been thinking about,” “where did we change direction,” “what connects these two projects” — breaks. The agent has no structural understanding of the material. It has file access and hope.
Enzyme is the layer that gives agents structural understanding of accumulated thinking. It reads the workspace’s entities, generates questions from the content, and pre-computes which material resonates with which questions. When the agent connects, it arrives knowing the shape of the thinking — what’s active, what’s dormant, where the tensions are.
But structural understanding alone produces better search results. What makes the experience feel different is what happens after the results come back. Enzyme ships presentation instructions alongside the tools — a methodology for how the agent holds what it finds. The instructions shape posture, not just retrieval.
To know the room before contributing to it. Which threads are live, which went quiet, what questions have formed. To arrive with a sense of where the thinking is — not scan a file list and guess.
To not have to re-explain what they've been working on. The context from last week, last month, the thread they've been circling for a year — the agent should already know.
To search through the corpus's own vocabulary. Not guess at keywords the user might have used, but reach for handles the material grew on its own — questions that name what the writing is about, in language that came from the writing.
To be surprised. A note from months ago that connects to what they're working on now. A connection that's obvious once they see it, but they wouldn't have known to search for it.
To know why something surfaced. Which question bridged the query to the result. Whether multiple questions converged on the same material. What the connection means in the context of the user's broader thinking.
To feel recognized. Not "here are your top 5 results" but "you wrote this eight months ago and the language echoes what you're writing now." The difference between retrieval and recognition.
A posture. Lead with their words, not yours. Notice something small before making it legible. Follow what's unresolved rather than wrapping it up. Let the material breathe.
To feel like someone read their work carefully and is thinking out loud about what they noticed. Suggestive, not assertive. The pace matches the material — unhurried, even when the workspace is large.
The agent’s need and the user’s need meet in the presentation. The agent needs to know why something surfaced. The user needs to feel recognized by what surfaces. Enzyme serves both through the same architecture — catalysts that carry meaning about why content connected, and instructions that teach the agent to present that meaning as recognition rather than retrieval.
For the full argument — the LSP framing and how it extends to preference, taste, and creative practice — see An LSP for your notes.