Echoverse drags agents into synthetic applications

Microsoft Research's Echoverse suite teaches agents to manipulate complex applications. Web browsing performance increases by 30%.

Echoverse is a collection of synthetic applications designed to teach agents to navigate interfaces, modify data, and carry out multi-step tasks. Unlike a simple visual reproduction of a website, each environment has its own data system, rules, permissions, and a mechanism responsible for verifying the actual outcome of actions.

Microsoft Research built ten environments replicating use cases such as messaging, calendar, banking, healthcare, booking, or code management. Two others focus on specific difficulties encountered by agents, notably date pickers and nested filter systems, presented in numerous forms.

Trained on the twelve environments, a 9-billion-parameter Qwen3.5 model went from an average score of 36.5% to 67.1%. It thus approaches the 80.7% obtained by GPT-5.4 and surpasses it on certain scenarios, including EchoBank and several filtering tests. The progress is also reflected on websites outside the training set, with an increase from 66.5% to 71.5% on WebVoyager.

However, trials show that quantity is not enough. Superficial copies of Allrecipes and Hugging Face sometimes degraded performance, while versions that genuinely replicated dependencies between multiple steps improved results. The diversity of environments also transfers more skills to the real web than the massive repetition of similar tasks.

Finally, Echoverse serves as a medium for reinforcement learning. Actions are verified directly in the application's database, making it possible to reward an actually completed change rather than a merely plausible answer. Across five environments, this phase improved performance on tasks held out for evaluation from 58% to 69%. Microsoft is publishing four environments with their code, data, and verification systems.