Echo Framework Turns Noisy AI Agent Logs into Training Data

A new framework called Echo, introduced in a preprint on arXiv (2605.21984), proposes a method to convert raw interaction logs from AI agents into structured training data. The approach addresses the limitations of static human-generated data, which is expensive to scale and constrained by creator knowledge. Echo leverages user refinement—the process by which users correct flawed agent proposals—as a primary feedback source. By filtering and structuring noisy, trial-and-error experience data, the framework aims to enable continuous model improvement from real-world deployments. The paper argues that widespread AI agent deployment provides low-cost access to massive experience streams, but raw logs are inefficient for direct training due to low information density. Echo operationalizes this transition by echoing environmental feedback back into the training loop.