Alibaba Synthesizes 64.9K Interactive Environments in Cowork-Gym (27B Beats 100B+ on ClawEval) & Harvard/MIT Open-Source Market-Based Agent Swarms (MPT)

Background and the Problem

Deploying autonomous agents into collaborative workflows exposes two major limitations: in human-agent coworking, users continuously refine requirements and directly edit files inside the shared workspace mid-task, breaking agents trained on static environments; in multi-agent swarms, training agents inside rigid, centralized orchestration graphs prevents individual skills and collective coordination from co-evolving.

Architecture and How It Works

Two new frameworks—Cowork-Gym (arXiv:2610.17670) and Market-Based Post-Training (MPT) (arXiv:2610.22322)—address human-agent coworking and decentralized swarm self-organization:

1. Cowork-Gym (64.9K Environments & Stateful Shared-Workspace User Simulator): Uses an execution-grounded build-execute-revise loop to synthesize 64,900 interactive coworking environments across 14 scenarios, pairing agents with a stateful user simulator that dynamically converses and directly edits shared workspace files;
2. Market-Based Post-Training (MPT) for Decentralized Swarms: Eliminates central coordinators by having each agent output a bid expressing its willingness to act at a given state; bids determine who executes actions and how collective task rewards are apportioned as local training credit.

Benchmarks and Measured Results

Extensive evaluations highlight substantial gains in adaptability and robustness:

• Cowork-Gym 27B Tops PawBench & ClawEval (Beating 100B+ Models): Post-training 9B, 27B, and 35B-A3B Qwen models with SFT, RL, and on-policy distillation yields average gains of +8.7, +6.4, and +4.1 points across six agentic benchmarks, with the 27B model achieving the #1 average score on both PawBench and ClawEval, surpassing models over 100B parameters;
• MPT Swarms Scale & Survive Agent Dropout: Across embodied crafting, navigation, and deep research, MPT swarms outperform fixed orchestration baselines, scale seamlessly to larger inference-time populations, and degrade gracefully under agent dropout.

Getting Started for Developers

MPT code and models are available at https://github.com/Embodied-Minds-Lab/MPT:

• Replace Central Routers with Learned Bidding: Use lightweight confidence bids in decentralized agent swarms to dynamically allocate subtasks and tolerate node failures;
• Handle Concurrent Human Edits in Coding Harnesses: Train and test IDE agents against concurrent workspace mutations so agents reconcile human file edits rather than overwriting them.