Greptile's TREX runs your code to find bugs static review misses

Greptile has added an execution layer to its AI code reviewer, called TREX, for Test, Run, Execute. Most AI review reads the diff and reasons about it. TREX instead runs the code. When the reviewer agent reads a pull request, it identifies issues worth investigating and spins up a separate TREX agent for each one, all in parallel, each inside a disposable sandbox that starts fresh in milliseconds and is thrown away when the run ends. To keep that fast without going stale, the system uses reusable base images and per-repository snapshots.

The point is to catch bugs that only appear at runtime: a logic error that needs a specific sequence of state, a UI regression that shows up after the page loads, a race condition that needs a real request in flight. Static analysis tends to miss these. Each finding comes with evidence rather than a guess, so a report can include screenshots, logs, API traces, and the execution script, and for an animation change TREX records a video of it running. The harness is model-agnostic, and the main agent and its subagents can even use different providers, so frontier models can be swapped in without rebuilding. Greptile says it deliberately trades latency for accuracy. You can read the writeup on the Greptile blog.

Why it matters

If your team relies on AI code review, the weak spot has been confidence: a reviewer that only reads code produces plausible comments you still have to verify by hand. A reviewer that runs the code and attaches a screenshot or a failing trace gives you something you can check in seconds, which changes how much you can trust it on the bugs that matter.