Wealthfront's engineering team has spent multiple years building a custom AI code review harness, after an engineer pasted a misbehaving branch diff into GPT-3.5 and asked "please review for bugs" — the response instantly caught a missed SQL filter in a refactor that had slipped past both automated testing and human code review.
The internal project started small: the team first asked "how many historical bugs could an LLM catch from just the diff?" The answer — many — kicked off a multi-year experiment to leverage AI in the Wealthfront code factory. The post then asks "can this new 'Claude Code' looping tool catch even non-obvious bugs through exploration?" The answer was yes, and the realization that other teams were working on similar problems but with different assumptions led the team to build an internal harness on top.
Per the engineering team's write-up, Wealthfront "didn't like the code review solutions available at the time" — general-purpose harnesses like Claude Code had too little structure, AI could not be trusted to stay on task or conserve its context window, and sub-agents (in their infancy) had problems with error propagation. "AI code review" companies were also targeting unrealistic goals. The team's approach — "build over buy" — reflects Wealthfront's belief that writing code is its core competency and that narrow, in-house solutions can outperform general-purpose vendor products.
The result, as described in the engineering blog, is a code-factory workflow that combines diff-level review for catching SQL-filter-style correctness bugs with sub-agent exploration across the codebase. In the team's own framing, this is described as their "first step toward building our ideal harness for AI-augmented code review," with deeper integration into the production-review pipeline expected to follow.
Details
- City
- New York
- Organization
- Wealthfront
- Continent
- North America
- Country
- United States