Agent Watch: The Week in AI Agents

Papers you should know about

This week: AI agents gained exciting new abilities that push them closer to true autonomy (although, let’s be honest, I could probably say this every week at the current rate of progress).

More specifically, researchers unveiled methods to plan more like strategists, remember like experienced assistants, reason across domains, and even shape group behavior through storytelling.

1. First, HyperTree Planning transforms how agents approach complex, multi-step tasks by enabling them to break goals into hierarchical sub-goals and explore multiple reasoning paths simultaneously. By allowing agents to self-organize their planning process without manual examples, this 3.6x performance improvement represents a crucial leap toward combining the knowledge depth of LLMs with the structured problem-solving of classical planning algorithms.

2. Next, Nemotron-CrossThink extends reinforcement learning beyond math into messier domains like law, physics, and history. By generating synthetic training data with built-in verification across diverse fields, this approach allows agents to improve through their own trial and error, achieving up to 30% higher accuracy on reasoning tasks while using 28% fewer tokens. It points toward a future where agents continuously improve through self-supervised practice.

3. MemEngine tackles the persistent "goldfish memory" problem by providing a unified, plug-and-play memory framework for agent systems. This standardized library dramatically lowers the barrier for equipping agents with persistent memory - enabling them to learn from past interactions and maintain context over extended periods, much like humans do with our working and long-term memory systems.

4. Finally, research on narrative alignment reveals how shared stories can shape multi-agent coordination. By priming agents with common narratives, researchers demonstrated remarkable improvements in cooperative behavior - suggesting a transparent, interpretable approach to aligning agents with human values and with each other through something as fundamentally human as storytelling.

These skills address core bottlenecks in agent design - from long-term planning and memory to general reasoning and multi-agent alignment - marking a meaningful shift in what autonomous AI systems can do.

Below we’ll go through the top developments of the week and why they matter, in an accessible deep dive for both curious newcomers and experts.