Echoverse: Training Agents in Deep Environments Boosts 9B Model Success Rate to 67%
burny_tech · x · 2026-08-03
Echoverse is a framework for training computer-use agents at scale using deep, evolving environments. Unlike shallow, self-contained web clones, it compiles specifications into stateful applications with real state, reset capabilities, and database-grounded grading.
Central to the framework is a co-evolution loop that reads every graded rollout twice: as repairs to the environment, tasks, and verifier, and as training signal for the model.
Experiments show that a 9B parameter agent trained on 12 such deep environments improves its average success rate from 36.5% to 67.1% across 14 evaluation splits, coming within 14 points of the much larger frontier model that taught it. In contrast, training on shallow environments hurts live-site accuracy (dropping from 80.0% to 75.0%). This suggests agents need environments with consequences, not just screenshots.
More from coding & agent
- Migrating from GPT-4o to GPT-5.1: Handling RAG Agent Response Style Regressions — IncreaseLocal2574 · 2026-08-03
- Using MCP to Seamlessly Invoke GlobalGPT in Codex for Writing and Video Generation — thetripathi58 · 2026-08-03
- Open-Source Skill Watches Videos and Generates Structured Notes — Roger_M_Taylor · 2026-08-03
- Java Dev Builds AI Agent in 30 Minutes Without Python Using LangChain4j — deepakatl1981 · 2026-08-03
- OpenRoboto Robotics Challenge: Fine-tune π0.5 to Win Token Rewards — const_reborn · 2026-08-03
- GitHub Project OpenCode: Orchestrating 7 Specialized Agents for Parallel Coding Tasks — tom_doerr · 2026-08-03