An AI agent dropped into an unfamiliar repository is functionally blind. It greps, it reads files, it guesses where things live. Every question - "where is this defined?", "what depends on this?" - costs a round-trip and a chunk of context.
The Genome removes the blindness.
What it is
The Genome is a live structural view of the entire codebase, held in context from the first message. Every file, every exported symbol with its signature, every dependency edge. It is ranked by importance, so the highest-impact files surface first.
It answers structural questions for free:
- *Where is
X?* → file and line, directly from the map. - *What does
Yexport?* → listed under the file. - *What depends on
Z?* → blast-radius count plus the list of importers.
No tool call. No file read. The answer is already in context.
Rebuilt after every edit
The map isn't a one-time scan. It's regenerated after every edit, so it never goes stale. When the agent renames a symbol or moves a function, the next thing it sees reflects that change - including the updated dependency edges.
Blast radius
Each file carries a blast-radius number: how many other files import it. Before touching a file that 40 others depend on, the agent knows to check the co-change graph and update the dependents too. It's the difference between a surgical fix and a broken build three files over.
Why a map beats search
Search tells you where a string appears. The Genome tells you what the code *is* - the symbols, the shapes, the relationships. Combine that with symbol-level editing and the agent stops behaving like a script running grep | sed and starts behaving like an engineer who already read the codebase.