June 17, 2026
Run code, stir chaos
TREX: An AI code reviewer that runs your code
This AI doesn’t just read your code — and commenters are already asking who’ll hack it first
TLDR: Greptile built an AI reviewer that runs software while checking it, aiming to catch problems that only appear when an app is actually used. Commenters are impressed by the idea but immediately split between excitement, security panic, and practical questions about whether it works in messy real-world setups.
Greptile’s new tool, TREX, is basically an AI code reviewer that actually runs the app instead of just staring at lines of text and pretending it understands everything. The pitch is simple: some bugs only show up when a program is doing something, not when it’s sitting still on a screen. So TREX spins up mini helper bots, tests specific problems, and can even return screenshots showing what broke. It’s a flashy upgrade to the classic “read the changes and hope for the best” approach — and yes, the company admits earlier versions were noisy, repetitive, and sometimes a little too confident about what they had really tested.
But the real show is in the comments, where the crowd instantly split into “finally, this is useful” and “cool, but how long until someone wrecks it?” camp. One of the loudest reactions was pure internet survival instinct: if you make an AI that runs code, people immediately want to know how fast it can be pwned. Others were more practical, cheering the idea because giving an AI a safe sandbox and letting it click around like a human could plug a huge hole in software testing. Then came the classic reality check: what happens when the code depends on outside stuff like cloud storage, secret passwords, and third-party services? And in perfect comment-thread fashion, one person ignored the AI drama entirely to praise the website’s design and that “chef’s kiss” sticky header. Priorities!
Key Points
- •The article presents TREX as Greptile’s AI code reviewer that executes code during pull-request review to detect runtime issues.
- •It argues that static code review, including most AI-assisted review, misses bugs that only appear when software runs, such as state-dependent logic errors, UI regressions, and race conditions.
- •TREX first existed as a standalone agent that generated and ran tests, but this produced irrelevant tests, missed edge cases, and duplicated work with Greptile’s main reviewer.
- •Greptile found that combining all review and execution work into a single agent overloaded context, so it shifted to a shared-context orchestrator model.
- •In the current design, the Greptile reviewer identifies issues in a diff and launches parallel TREX subagents that investigate specific problems, set up environments, and return evidence such as screenshots.