Agents Learn to Code (Themselves): Self-Improvement, Code-As-Infrastructure, and the End of Manual Pipelines

Tools

Unitree SDK2 Drops: Build Real Robotic Agents with Native C++ and Python APIs

This isn't a simulation. Unitree's open SDK gives you direct control over physical robot agents, from motion planning to sensor fusion. The robotics-agent integration layer just got a real foundation.

CrewAI Deepens Agent Memory: Long-Term Recall and Contextual Learning in Production

CrewAI's memory system now supports persistent, structured recall across tasks and sessions. This is how you stop agents from being goldfish—context that actually sticks.

LangSmith Adds Agent Tracing and Evaluation: Debug Multi-Step Agent Flows in Real Time

When your agent chain breaks, you need to see where and why. LangSmith's tracing gives you visibility into token usage, latency, and decision points across agent runs.

Research

Self-Improving Agents: New Research on Autonomous Code Generation and Tool Creation

The paper explores agents that generate and refine their own Python tools mid-task. No more hand-coded pipelines—the agent iterates on its own infrastructure.