What's New in Explyt 5.3
Model Selection in Settings for Individual Users
Now individual users can choose the model they will use with the agent.
Agent Modes (Modes)
We have added several modes of operation for the agent.
Agent - a full-fledged agent that can read and edit files, identical to the old default chat mode. It is best suited for day-to-day work not related to project exploration and testing.
Ask - read-only agent mode: it can read files in the project, read pages on the internet. It is best suited for project exploration tasks (e.g., onboarding into a new project), brainstorming implementation ideas, analyzing the causes of bugs.
Tester - mode focused on test generation.
Although you can generate tests in Agent mode, this mode will provide better results.
For example, on our internal benchmark of 33 real projects (with endpoints of ~2000 lines, ≥150 KLOC with Spring, TestContainers, Java, Kotlin) with all other things being equal (same model and same input prompt) the Tester agent gave significantly better results than the basic Agent.
| Test generation task from source code | Agent | Tester |
|---|---|---|
| Percentage of compiled test classes | 76% | 94% |
| Percentage of passing tests | 67% | 90% |
| Test coverage | 64% | 80% |
The Tester mode is currently only available in IDEA (and its forks) for Java and Kotlin.
Smooth Chat Display
We improved the UI by making the display of AI-generated text smooth so it's less irritating to the eyes.
Agent Edit Tool Fix
As part of our continuous improvement of our AI agent, we have implemented significant enhancements to the file editing tool (edit_tool) to increase its reliability and success rate. We want to share with you benchmark results demonstrating a noticeable improvement in agent effectiveness.
Technical benchmark results show statistically significant improvement in key metrics. File editing success rate increased by 8.4% in small tasks (from 82.47% to 89.47%), by 4.86% in complex tasks (from 92.54% to 97.06%), and by 6.42% in test generation (from 90.55% to 96.36%). The agent has also become more reliable in following requirements and violates explicit user instructions less frequently.
For developers and testers, this means more predictable and safe interaction with the agent. The agent now makes fewer unintended changes, causes less code damage, and more frequently completes tasks successfully. The improved reliability of the editing tool reduces time spent debugging and fixing errors introduced by the agent.