AI Agents of the Week: Papers You Should Know About

Get ahead of the curve with LLM Watch

In this week’s agent highlights, researchers introduced new frameworks to orchestrate complex multi-agent workflows with greater oversight, addressed the long-term memory problem in LLM-based agents, and demonstrated new methods for adapting agents to dynamic environments without retraining. Major themes include:

• Enhanced Planning & Orchestration: Two new frameworks (e.g. Alpha Berkeley and SOAN) tackle the coordination of complex tasks. They introduce structured planning approaches – from plan-first execution with optional human approval to self-organizing agent networks – aimed at making multi-step autonomous workflows more scalable, transparent, and robust.

• Memory Systems: Building on August’s focus on memory, a new Multiple Memory Systems approach draws inspiration from cognitive psychology to dramatically improve how agents store and retrieve long-term knowledge, leading to more coherent and context-aware interactions.

• Reliability & Self-Correction: A “cognitive operating system” framework (COCO) injects continuous oversight into agent teams. By monitoring for errors and rolling back or correcting them on the fly, it prevents small mistakes from snowballing – boosting overall success rates and establishing a new state-of-the-art in workflow reliability.

• Adaptability in Dynamic Environments: A novel paradigm (TAPA) shows that large language models can help agents adapt on the fly to changing conditions (like cyber-attacks or shifting team goals) by synthesizing new action strategies as needed This hints at a future where agents continuously evolve their own “skills” without expensive retraining.

In the breakdowns below, we dive into five standout papers, explaining their core innovations, why they matter for autonomous agents, what problems they solve, and what possibilities they unlock going forward.