title: "Tsinghua and Moonshot open-source AgentENV platform" slug: "tsinghua-and-moonshot-open-source-agentenv-platform" published: "2026-08-09" beat: "Tools" tags: ["Tools", "Business"] creator: "Agentry Newsroom" editor: "Susanne Sperling, Editor — Human in the Loop" tools: ["Claude (Anthropic)", "Perplexity Sonar"] creativeWorkStatus: "verified" dateReviewed: "2026-08-09" aiActArticle50: "compliant" humanView: "https://agentry.news/tools/tsinghua-and-moonshot-open-source-agentenv-platform" agentView: "https://agentry.news/agent/tsinghua-and-moonshot-open-source-agentenv-platform"
Tsinghua University's MADSys Lab and Moonshot AI jointly open-sourced AgentENV, a distributed execution platform for large-scale agentic reinforcement learning, on July 25, 2026. The infrastructure po
Drafted by an AI agent. Verified by Susanne Sperling, Editor — Human in the Loop. AI policy.
Tsinghua University's MADSys Lab and Moonshot AI jointly open-sourced AgentENV (AENV), an agent execution infrastructure platform designed for large-scale agentic reinforcement learning, according to an announcement on the kvcache.ai blog kvcache.ai. The release, which occurred on July 25, 2026, marks a significant shift toward making production-scale agent training infrastructure publicly available.
AgentENV is described as "a distributed platform for running agent environments at scale" GitHub. The system was built to handle the computational demands of training agentic RL systems at the scale required for production language models. Moonshot's Kimi account confirmed that AgentENV components "power agentic RL training for Kimi K3" and support workflow features including snapshot, resume, and fork capabilities X/Twitter.
The platform addresses a concrete technical gap: while agentic RL has emerged as a training paradigm for autonomous AI systems, the infrastructure for running thousands of parallel agent-environment interactions—collecting training data, rolling back failed episodes, and resuming from checkpoints—has remained either proprietary or unavailable at scale. By open-sourcing this layer, Tsinghua and Moonshot are enabling other teams to build and benchmark agentic systems without building custom distributed execution stacks.
The project is hosted under the kvcache-ai GitHub organization, established as a joint initiative between Tsinghua's MADSys Lab and Moonshot AI GitHub. Moonshot's announcement noted the release occurred "in collaboration with kvcache-ai," and identified Tsinghua's team as the initiator of the joint open-source organization kvcache.ai.
The timing reflects broader momentum in the agent economy toward open infrastructure. As agentic capabilities move from research demonstrations to production deployments, the software layer supporting agent training and execution has become a critical resource. By releasing AgentENV, the partners are positioning themselves within the developer ecosystem while establishing reference standards for how large-scale agent environments should be architected.
Developers can now access the full codebase, documentation, and examples on GitHub. The v0.1.0 release includes the core distributed system components used in Kimi K3 training, though the project is likely to evolve as external contributors engage with the platform.