Alibaba Qwen Team Launches AgentWorld to Simulate Multi-Agent Environments

Alibaba’s Qwen team officially introduced Qwen-AgentWorld on June 22, 2026, with wider public announcements following on June 24, 2026, launching a dual-model system capable of simulating seven different digital environments simultaneously. Unlike typical AI agents designed to execute tasks, this system functions as a “language world model” that mimics how computer environments respond to user actions. By replacing unpredictable live systems with structured, simulated training loops, the software aims to produce more autonomous and reliable AI agents.

How Language World Models Flip Traditional Agent Training

Most AI agents operate in a reactive loop: they observe a system state, select an action, and wait for the system’s output. Qwen-AgentWorld reverses this execution logic. Instead of training an agent to select the next action, the system is optimized to predict what the environment will return after an action is taken.

The model operates as a text-based simulator that maps action inputs to anticipated feedback. For example, if an agent inputs a command into a command-line interface, Qwen-AgentWorld predicts the exact string output, down to byte counts, file formats, and directory structures. If the agent triggers a database tool, the system estimates the exact JSON payload the tool would return.

This predictive approach is intended to replace slow, fragile live connections. Instead of communicating with real databases or running sandboxed command-line terminals, developers can run training pipelines entirely within the simulated world.

Seven Simulated Domains in a Unified Architecture

The release unifies seven major operational environments into a single model architecture. Rather than running separate specialized simulators, a single instance can run transitions across both text-based and graphical user interface (GUI) formats.

The seven simulated domains consist of:

• Model Context Protocol (MCP): The standard interface protocol that lets AI applications talk to external data sources and local tools.
• Search: A simulated search engine environment that mimics web queries under controlled parameters.
• Terminal: A mock Linux command-line environment that replicates system responses.
• Software Engineering (SWE): A software development environment capable of handling repository structures and code execution outputs.
• Web: A browser environment that simulates user navigation and DOM changes.
• Desktop OS: A desktop operating system simulator.
• Android: A mobile operating system simulator.

For GUI-based environments like Android, Web, and Desktop OS, the models do not process visual pixel arrays. Instead, they interpret and generate textual accessibility trees and structured user-interface view hierarchies. This choice keeps the environment simulation computationally lighter while maintaining detailed state transitions.

Mixture-of-Experts Architecture and Specifications

Alibaba developed two separate Mixture-of-Experts (MoE) model sizes for this release. Both variants feature a 256K context window (specifically 262,144 tokens), allowing them to keep long interaction histories in memory:

• Qwen-AgentWorld-35B-A3B: A model containing 35 billion total parameters, which activates 3 billion parameters per token during execution. Its weights are open-source and released under the Apache 2.0 license.
• Qwen-AgentWorld-397B-A17B: The larger flagship model with 397 billion total parameters, activating 17 billion parameters per token. Alibaba has kept the weights of this variant proprietary.

The default sampling parameters recommended by the developers for model inference include:

• Temperature: 0.6
• Top_p: 0.95
• Top_k: 20
• Max Output Length: 32,768 tokens to accommodate long multi-step trajectories and detailed environment observations.

The Three-Stage Training Pipeline

To teach these models how computers behave, the Qwen team trained them on more than 10 million environment interaction trajectories recorded from real agent runs. This data passed through a three-stage training process:

1. Continual Pre-Training (CPT): Teaches the model basic transition dynamics, file structures, and API responses. The goal is to inject broad world knowledge by combining environment interaction traces with specialized corpora from domains like cybersecurity, medicine, and industrial control.
2. Supervised Fine-Tuning (SFT): Activates multi-step next-state-prediction reasoning. During SFT, the model is trained to generate explicit reasoning chains—often called “Chain-of-Thought”—inside <think> blocks before emitting its final environment prediction.
3. Reinforcement Learning (RL): Applies Group Sequence Policy Optimization (GSPO) to sharpen simulation fidelity. Because evaluating a simulator’s accuracy is difficult, Alibaba avoided standard Turing-test style discriminators. Instead, they combined strict rule-based verifiers with an LLM-as-a-judge rubric that scores the output across five dimensions: format, factuality, consistency, realism, and quality.

Practical Applications in Simulation Control and Agent Warm-Up

The Qwen team’s research details two distinct methods for using these world models to improve downstream agent training:

• Decoupled Simulation (Sim RL): The world model acts as an external simulator, replacing the real-world environment during the agent’s reinforcement learning loops. A major bottleneck in agent development is the fragility of real-world training environments. Live search engines or actual containerized terminals cannot easily be forced to simulate rare or adversarial conditions. Because Qwen-AgentWorld is entirely virtual, developers can force the environment to inject artificial errors, report low disk space, or return corrupted API packets on demand.
• Unified Agent Foundation Model (Warm-Up): The world model training objective is used as a direct warm-up phase for the agent itself. By first training an agent model to predict the next state of its environment, the model develops an internal understanding of system dynamics. When the agent is later fine-tuned on actual tasks, this background knowledge yields immediate benefits.

In testing, this warm-up phase increased downstream agent performance by an average of nearly 9 points across seven benchmarks. On completely out-of-domain benchmarks that were absent from the training set, performance jumped by 11.28 points on Claw-Eval and 8.96 points on BFCL v4.

Additionally, during testing of the standalone simulator (Sim RL), training an agent inside 4,000 “OpenClaw” environments that were entirely out-of-distribution for the model yielded zero-shot environment generalization, resulting in performance gains of 4.3 points on Claw-Eval and 7.1 points on QwenClawBench.

The evaluation process for testing the simulation quality follows a three-step command sequence using the AgentWorldBench dataset:

# Step 1: Run world model inference
python eval.py infer \
  --data-dir ../AgentWorldBench \
  --model-base-url http://localhost:8000/v1 \
  --model-name Qwen/Qwen-AgentWorld-35B-A3B \
  --output-dir ./results

# Step 2: Run LLM judge scoring
export OPENAI_API_KEY="your-api-key"
python eval.py judge \
  --predictions ./results/predictions.jsonl \
  --judge-base-url https://api.openai.com/v1 \
  --judge-model gpt-5.2-2025-12-11 \
  --output-dir ./results

# Step 3: Aggregate and display scores
python eval.py score --predictions ./results/judged.jsonl

Benchmark Results on AgentWorldBench

Alongside the models, the researchers released AgentWorldBench, a new evaluation suite built to measure how closely a simulation matches real-world outcomes. The benchmark compiles ground-truth interactions of five frontier models tested across nine established industry datasets.

The flagship Qwen-AgentWorld-397B-A17B achieved an overall score of 58.71 on the benchmark. For comparison, the paper reports that GPT-5.4 scored 58.25 and Claude Opus 4.8 scored 56.59 under the same evaluation parameters.

The largest performance margins appeared in text-heavy environments:

• Terminal: The flagship model scored 57.73, compared to GPT-5.4’s 53.69.
• Software Engineering (SWE): The model reached a score of 68.49, compared to 66.29 for GPT-5.4.

GUI-based environments like Android showed much tighter margins, which the Qwen team attributes to the visual pre-training advantages of proprietary models.