Modern web applications often suffer from bugs that are difficult for automated agents to resolve because those agents lack deep context. When an AI agent is tasked with fixing a bug based only on surface-level observations—such as a screen recording, a network log, or a console error—it must often guess the underlying cause. If the guess is incorrect, the agent may enter a cycle of failed repairs, potentially making the issue worse. Replay, a company focused on observability, argues that the missing ingredient for reliable automated bug fixing is "perfect observability": a way for agents to see every statement executed, every state change, and every data dependency without loss of detail.
To provide this context, Replay utilizes a time-travel debugger for Chromium and Node.js. This technology captures non-deterministic inputs with low overhead, allowing the runtime to be replayed deterministically. By recording these interactions, the system can unpack the execution into a database where every point of the application's behavior can be queried. This allows an AI agent to work backward from a failure to the exact root cause. For example, while a standard QA tool might identify a "dead button" and provide a reproduction script, Replay QA can identify the specific missing `onClick` handler on a React component and provide the evidence needed for an agent to verify and fix the code directly.
In a recent experiment, Replay QA was run on 40 applications promoted by major app-building tools, resulting in the identification of over 900 bugs. The system handles complex issues like layout shifts by providing reports that include contributing factors such as specific React commits, network requests, and variable values in render functions. However, while Replay QA provides the "what" and "why" of a bug, the actual repair is still performed by an agent or a "software factory" that converts detected problems into durable test cases. Furthermore, some developers argue that existing logs and reproduction steps may already provide sufficient context for capable agents, suggesting that the need for a full time-travel debugger depends on the sophistication of the agent being used.",
This image serves as the primary visual for the Replay QA blog post discussing their methodology for automated bug fixing.