Blitzy’s autonomous coding bet: Every codebase is already a graph

Blitzy is making a simple argument with big consequences: autonomous coding won’t scale until agents understand the codebase as a connected system, not a pile of files. That idea sits behind the company’s use of Neo4j and its push to run thousands of coding agents in parallel after raising $200 million at a $1.4 billion valuation in May.

Neeraj Deshmukh, Blitzy’s director of engineering, said the real problem isn’t whether AI can write code. It’s whether it understands the system it’s changing. A small edit can ripple far beyond the line in front of it, which is exactly why Blitzy starts by reverse-engineering a customer environment and building a live graph of the codebase.

That graph is tied into systems like GitHub and GitLab, so it updates as people or agents make changes. Deshmukh argued that this matches how software is already built. Modules, files, functions, objects, classes and variables all connect to one another. In his view, the codebase is a graph before AI ever enters the picture.

The alternative is the usual grab bag of vector search or grep, which can chew through an agent’s working memory fast. Deshmukh put effective context at about 200,000 to 300,000 tokens, or roughly 20,000 to 30,000 lines of code. At a 100-million-line codebase, that gets thin quickly. A graph gives the agent a tighter, more deliberate slice of what matters, instead of a fog of nearby text.

Blitzy says that structure lets it take on whole projects rather than breaking them into the familiar sprint machinery of epics, user stories and tasks. Humans still approve an Agent Action Plan before coding begins, and every line gets tested immediately. The company also pointed to an 84.95% score on SWE-Bench Pro in June. On top of that, other agents watch the ones doing the work, while Cypher queries in Neo4j act as a guardrail: if the query is wrong, it returns nothing.