CoworkEnv: Synthesizing Training Environments for Long-Horizon Cowork Agents
We present CoworkEnv, a capability-guided framework for synthesizing executable training environments for AI coworkers.
ProofPaper ↗
Key points
- AI coworkers must coordinate tools, communicate across turns, and act on changing workspace state while respecting workflow constraints.
- An SFT-only extension to ClawMark demonstrates applicability across coworker benchmarks.
- We further compare RL training dynamics across configurations that vary reward processing and credit assignment.
- Improving training rewards provide preliminary evidence that the constructed environments support policy optimization.
Sources (2)
- [1]CoworkEnv: Synthesizing Training Environments for Long-Horizon Cowork AgentsHugging Face Daily Papers · Oct 18, 12:00 AM
We present CoworkEnv, a capability-guided framework for synthesizing executable training environments for AI coworkers.
AI coworkers must coordinate tools, communicate across turns, and act on changing workspace state while respecting workflow constraints.
- [2]CoworkEnv: Synthesizing Training Environments for Long-Horizon Cowork AgentsarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 18, 02:20 AM · same content
Extractive summary: sentences quoted from the sources.
Before this
- Oct 17, 2026Learning to Program Retrievals: Composing Search Primitives via Reinforcement Learning
- Oct 17, 2026HybridResearcher: Bridging the Structured-Unstructured Boundary for Deep Research Agents
- Oct 16, 2026Pronouncing Hard Words: Phoneme Adaptation of a Pretrained Conversational Text to Speech Model
- Oct 16, 2026Welcome Beam: a 501B open model you can own, customize, and scale
- Oct 16, 2026Tree-Aware Posterior Distillation for Long-Horizon Agentic Reinforcement Learning
- Oct 15, 2026DeciCAD: Harness-Grounded Learning of Action Policies for CAD Agents