PostHog SQL parser hits 70x speedup

This is a landmark case study showing that LLMs can write complex, high-performance systems code when paired with rigorous testing pipelines. It shifts the agentic coding narrative from simple scripting to structured, test-driven synthesis.

• –Replacing ANTLR's runtime graph-walking interpreter with native, hand-rolled Rust code yielded a massive 454x average speedup on production queries.
• –The developer's primary role shifted from writing code to building validation loops, utilizing Hypothesis for property-based testing to compare AST outputs with the old parser.
• –Brittle lookahead fixes by the LLM were solved through targeted prompt engineering, forcing the agent to load relevant grammar rules into context before editing.
• –The final result is a highly complex 16,000-line Rust parser (using a Pratt expression parser and speculative backtracking) written entirely by Claude.
• –This highlights a new development pattern: using parser generators to establish correctness (as an oracle) and letting LLMs generate the high-performance implementation.