June 2026 Summaries

Shlok, a software engineer at Greptile, discusses the development of TREX, a novel code reviewer that not only examines pull requests but also runs the code to identify runtime issues, addressing limitations of traditional static code reviews. Inspired by Michael Fagan's 1976 introduction of formal code inspection at IBM, TREX was created to identify bugs that appear during code execution, such as logic errors and race conditions, which static reviews often miss. Initially developed as a separate product, TREX faced challenges with test generation and context sharing, leading to its integration with Greptile's main reviewer to better manage context and orchestrate dedicated agents for specific issues. This integration allows TREX to leverage various modalities, including screenshots, logs, and videos, to provide comprehensive and trustworthy artifact sets for each finding. The system is model-agnostic, enabling the use of multiple models across reviews without quality penalties, focusing on infrastructure, artifact generation, and evaluation rather than model intelligence. TREX utilizes sandboxed environments for reliable execution, ensuring that each review is a reproducible experiment with verifiable results, ultimately aiming to transform code review into an automated validation suite to eliminate bugs entirely.

Greptile has introduced TREX, an execution layer for code review that enhances its ability to detect bugs by running the code in addition to reading it. Addressing the limitation of static code review, which can miss issues only apparent during execution, TREX (short for Test, Run, Execute) identifies and runs relevant code behaviors in a sandbox environment to detect problems such as broken page renders or endpoint errors. This approach results in a 20% increase in bug detection during evaluations. TREX provides comprehensive evidence, including logs and screenshots, to substantiate its findings, allowing for more thorough validation and a step toward achieving software with no bugs.

Greptile is an AI agent that enhances the pull request review process for developers by providing context-aware feedback, akin to that of a senior engineer, which helps catch bugs before they are shipped and reduces the reliance on human reviewers. This system leverages a multi-model architecture where various AI models are assigned specific tasks, such as classifying developer replies or indexing codebases, to optimize for speed, cost, and accuracy. A recent evaluation of the NVIDIA Nemotron 3 Ultra demonstrated its efficiency and cost-effectiveness, outperforming existing models in tasks like reply classification and codebase indexing, thus enabling Greptile to deliver faster and more accurate feedback. This approach allows Greptile to remain affordable and scalable, serving 22,000 engineering teams by continuously updating its models to maintain high-quality reviews and accurate codebase mappings.