Clone
Rebuild frontend, backend, database, and local tooling.
VeriEnv
Verifiable environments for self-evolving web agents.
VeriEnv automatically clones real-world websites into fully executable synthetic environments, exposing internal state through a Python SDK so agents can learn from deterministic, verifiable rewards instead of brittle LLM-as-a-judge feedback.
- 0 synthetic websites
- 0 verifiable tasks
- +0.00 / +0.00 WebArena gains over base models
Safe Exploration
Clone first, train later
Learn in controlled website replicas without touching live production environments.
Verified Rewards
Tasks + judges grounded in code and DB state
Every trajectory can be checked programmatically through the synthetic environment.
The core idea is simple: treat language models as environment creators, not just action policies. By reconstructing websites into instrumented training worlds, VeriEnv makes web-agent self-evolution safe, repeatable, and scalable.
Motivation
Why real websites are a bad training substrate
Direct self-evolution on the open web is unsafe, hard to reset, and often judged by ambiguous instructions or non-verifiable reward signals. VeriEnv replaces that loop with controllable, executable replicas.
Traditional self-evolution vs. verifiable environments
The paper’s motivation figure contrasts fragile real-world exploration with VeriEnv’s synthetic websites, validated tasks, and deterministic reward signals.
Open motivation figureMethod Overview
From a real website to a verifiable training environment
VeriEnv uses a coding agent to reconstruct the full stack of a website, then generates tasks and judges that interact with both the UI and the database through an SDK for end-to-end verification.
Generate
Create tasks at multiple difficulty levels automatically.
Verify
Judge outcomes deterministically with executable checks over environment state.
Instrumented website cloning pipeline
This overview figure shows the full VeriEnv loop: clone a website, expose code/database interfaces, generate task-judge pairs, and train agents using verified reward signals.
Open method overview figureWebsite Showcase
Examples of websites reconstructed inside VeriEnv
These snapshots come from synthetic sites built for agent training. The hero animation cycles across multiple reconstructed websites to emphasize scale and diversity.
Allrecipes
Recipe discovery and structured cooking workflows.
Budget
Task-heavy booking flows with bold transactional UI.
Tesla
Highly visual commercial hero sections and product landing flows.
Healthline
Editorial card layouts and article-heavy browsing patterns.
CA.gov
Public-sector portals with utility-first navigation.
Eventbrite
Event discovery pages with strong promotional hero design.
Why It Matters
Environment scaling instead of prompt hacking
VeriEnv shifts the bottleneck for web-agent training from unverifiable language supervision to scalable environment construction. More websites mean more tasks, more coverage, and more stable reinforcement signals.
Takeaway
Train web agents in environments you can trust.
VeriEnv makes it possible to scale self-evolving web agents with synthetic websites that are faithful enough to be useful and instrumented enough to be verifiable.