The Week in AI Agents: Papers You Should Know About

From ReAct to Pre-Act and Strategy-Augmented AI Planning

This week: Enhancements to AI agents’ core capabilities in planning, memory, decision-making, and coordination. The focus is shifting from autonomy towards more resilient agentic systems. Recent publications have introduced new methods that enable AI agents to plan complex multi-step actions with greater reliability, maintain structured memories over long tasks, coordinate and cooperate in multi-agent settings, and even adapt strategies against adversaries.

From large language model (LLM) agents that explicitly reason through step-by-step plans (dramatically improving task success rates) to importing concepts like altruism from Biopsychology to improve multi-agent cooperation.

Each of the works highlighted in this article not only solves a key technical challenge (such as context-length limits or brittle planning) but also provides a vision of how future AI agents can achieve higher-level goals more safely and efficiently:

1. Pre-Act: A technique that improves agent performance by generating multi-step plans before action

2. Strategy-Augmented Planning (SAP): A framework for LLM agents to model and exploit opponent strategies in adversarial settings

3. Hamilton's Rule: A biological principle adapted to enable altruistic behavior in multi-agent systems

4. Community-Based Learning: A method for organizing agents into overlapping communities for more efficient learning

5. RedTeamLLM: An architecture for autonomous cybersecurity testing with advanced planning and memory capabilities

So let’s take a closer look at these papers, explaining their innovations, the problems they tackle, the techniques they employ, and why these advances matter for the future of intelligent autonomous systems. Enjoy!