AI Papers Reader

Personalized digests of latest AI research

View on GitHub

Synthetic Workspaces Aim to Train AI Agents for Changing Work

Language-model agents are increasingly asked to complete multi-step jobs in shared workspaces, reading files, running tools, and revising documents, while the people who assigned the work change their requirements or swap out inputs partway through. Training agents for this kind of “cowork” requires practice environments that align the task goal, files, tools, simulated user behavior, and success checks. Building them by hand is slow and error-prone. A new paper from Alibaba’s Token Foundry proposes generating such environments automatically, at scale.

The system, called Cowork-Gym, builds each environment through a loop of building, executing, and revising. A builder agent assembles an initial workspace, the available tools, and a simulated user. A separate verifier constructor writes checks that score the outcome, combining deterministic code tests with a language-model grader. An issue detector then runs the environment, confirms that the checks separate valid completions from incomplete or malformed work, and routes defects back to the component responsible. Only environments that pass are used to collect agent trajectories, the step-by-step records of an agent’s actions, which are then filtered for validity.

A central feature is a “stateful” user simulator, meaning it tracks the task’s progress and can change the shared workspace, not just send messages. The paper’s example involves a sales report. An agent analyzes a file called sales.csv and drafts report.md. The simulated user then replaces sales.csv with corrected data and asks the agent to exclude refunds, so the agent must recompute its totals and revise the report. The authors argue that such mid-task changes test whether an agent can revisit earlier work, rather than simply completing a fixed request.

The team synthesized 64,888 environments across 14 application scenarios. They used them to post-train three Qwen models, Qwen3.5-9B, Qwen3.6-27B, and Qwen3.6-35B-A3B, through supervised fine-tuning and reinforcement learning, with on-policy distillation as an option. Averaged across six agentic benchmarks, the 9B, 27B, and 35B-A3B versions gained 8.7, 6.4, and 4.1 points over their base checkpoints, improving in 17 of 18 model–benchmark comparisons. These are percentage points on the averaged score. Relative to its base model, the 9B gain is roughly 24 percent. The authors’ 27B model averaged 0.6119, placing second behind GLM-5.2 (0.6157). It recorded the highest scores among the evaluated models on PawBench (0.8400) and ClawEval (0.8508), a group that includes models with more than 100 billion total parameters.

The gains are uneven. On WildClawBench, the 9B model improved by only 1.36 points, the 35B-A3B model declined by 0.52 points, and all three of the authors’ models remained substantially behind GLM-5.2. The verifiers also depend partly on language-model graders, which the authors acknowledge can be biased or overly permissive. They say larger-scale human audits are needed to measure false-acceptance and false-rejection rates. The environments, the authors add, are synthetic approximations of personal work that emphasize browser, file, data, and office tasks.

The authors frame the work as a step toward agents that improve as their training environments expand. The reported results come from their chosen benchmarks and do not by themselves show how these agents would perform on real users’ workloads.