🖥️ Executable training environments for coding agents are far scarcer than the trajectories we already have — this paper flips that imbalance around. Title: Terminal-Universe: Turning Agent Trajectories into Scalable Terminal Environments URL: The framework reads the file-operation history embedded in existing agent trajectories, reconstructs executable, verifiable environments from them, and then synthesizes large batches of new training tasks on top. Key points 🔄 Reconstructing environments from trajectories It replays the read/write/edit operations in a trajectory to rebuild a partial workspace, then a completion agent fills in missing files and dependencies. A read-only verifier filters out workspaces that aren't actually task-sufficient. 🌐 Expanding tasks along width and depth Cross-WS synthesis links dependencies across multiple repos for "width," while Multi-Round dialogue adds user feedback for "depth." Multi-Round conversations average 4.51 rounds, with 69.6% following a fail-then-fix pattern. 📈 Double-digit performance gains Fine-tuning Qwen3.5-27B on this data lifts Terminal-Bench 2.1 by +11.9pt and EvoCode-Bench v2 by +13.8pt, beating every comparably-sized baseline it was tested against. It's a grounded way to scale agent training data by turning an already-abundant resource — leftover trajectories — into reusable, verifiable environments. #AgentTraining# #LLM#

🧩 Agent training environments are often surprisingly idealized: tools always respond cleanly, users always give clear instructions. But how well does an agent raised on that assumption hold up against real timeouts, errors, and sudden mid-task corrections from a real user? That's the question this paper takes on. The authors mined over 24,000 real tool failures across 557 MCP servers, distilling them into 8 failure families and 1,170 "recovery skills." They then teach an agent to reflect from three angles - recovering right after a failure, replanning from scratch, and responding to natural-language corrections from a user - and sharpen that ability further with reinforcement learning on a mix of noisy and clean rollouts. That's Agent-Reflex. The results back it up: average accuracy across 5 long-horizon tool-use benchmarks jumps from 31.6% for the base model to 58.6%, and even under injected noise the model holds at 45.9%. Baking the ability to recover from failure into the model's parameters, rather than relying on prompting, is what makes the difference. Title: Agent-Reflex: Teaching Language Agents to Act Reflectively URL: #LLMAgents# #ReinforcementLearning#

🔔 AI Agent Bootcamp | Lesson 2 Goes Live Tomorrow at 8 PM In Lesson 1, we started from the fundamentals of large language models and explored how AI learns to understand context. Tomorrow, we’re taking it one step further 👇 📚 Lesson 2 | How Do LLMs Learn to Understand Instructions and Generate Language? We’ll dive into attention mechanisms, instruction alignment, MoE, and the differences between models to better understand: · Why can AI understand and follow your instructions? · Why can the same prompt produce very different results across different models? This is only step two of our 10-week hands-on AI Agent training journey. Coming up next: Prompt → Workflow → RAG → Agent → Real-World AI Projects If you want to systematically learn how to build with AI Agents, add the TinTinAI Assistant to learn more about the bootcamp 👇 #TinTinAI# #TinTinLand#

Just wrapped up my first private Hermes Agent training class. 2 hours of going deep on Hermes, agents, workflows, tooling, and just how POWERFUL this agent really is. This is exactly the kind of work I want to be doing more of. Helping people understand what’s possible with AI agents, then showing them how to turn that into something useful in the real world. Huge thanks to @CMichaelGibbs and @xturnlgod for trusting me with the session. This was a big milestone for me. More to come. 🚀

📱 You want to train an agent that operates a phone. But real devices are costly, hard to reset, and hard to auto-verify, while mock environments alone are too simplistic to transfer to real phones. That dilemma has long blocked phone-agent training. The PhoneBuddy team refused the either-or and combined both. They first give Qwen3.5-4B a shared foundation with supervised fine-tuning over 950k action steps from real and mock environments. Then they blend real-app reinforcement learning fifty-fifty with PhoneWorld, a mock environment that reconstructs runnable Android apps from real GUI traces. The mock side auto-checks success with rule-based verifiers, while the real side has Gemini 3.1 generate rubrics and a large model score the trajectories, a clean division of labor. The payoff: average success on 150 real-phone tasks rose from 36.67% to 45.33%, and AndroidWorld climbed from 60.3% to 83.2% across stages, with this tiny 4B open model beating Gemini 3.1 Pro on that benchmark. The authors show mock and real are complementary rather than substitutable, while honestly noting that cross-app workflows remain the biggest open challenge. Title: PhoneBuddy: Training Open Models for Agentic Phone Use URL: #PhoneAgents# #ReinforcementLearning#

🔥DeepSeek has published a new paper detailing its Agent training infrastructure, DSec, co-authored by Liang Wenfeng. DeepSeek Elastic Compute is a production sandbox platform that exposes FnCall, container, microVM, and full-VM sandbox backends through a unified SDK. A single production-scale DSec unit spans ~160 nodes, serving ~3M sandboxes per day. In production, it supports 380K+ concurrent sandboxes and sustains 5,000+ sandbox creations per second.

Are brands keeping up with consumer expectations? Our latest blog shares 4 key CX lessons from CRS Scottsdale, covering AI, customer loyalty, agent training, and more. Read the insights and what they mean for CX leaders: #ExecsInTheKnow# #Research#

Reflection just launched Beam which is a 501B open weight model for coding, reasoning and agents that it says can match GLM-5.2 while using ~4x less inference compute. The more interesting part is RL used 10,500 $NVDA GB300s versus 6,144 for pretraining and generated 100M+ rollouts showing agent training is becoming a huge inference workload. That helps explain why Reflection locked in $7B+ of compute across $SPCX and $NBIS through 2029 while positioning Beam as a U.S. open model for enterprises and sovereign AI.

We believe instant generation is where video models are headed. Last week, we released Solaris and GWM Worlds 2, and detailed our approach to agent training in digital and physical worlds. Today, we’re sharing a broader look at our research efforts in real-time video generation. Read more at the link below.